Upload
others
View
8
Download
0
Embed Size (px)
Citation preview
This article was downloaded by: [138.67.42.201] On: 07 November 2019, At: 11:10Publisher: Institute for Operations Research and the Management Sciences (INFORMS)INFORMS is located in Maryland, USA
Operations Research
Publication details, including instructions for authors and subscription information:http://pubsonline.informs.org
Hierarchical Benders Decomposition for Open-Pit MineBlock SequencingThomas W. M. Vossen, R. Kevin Wood, Alexandra M. Newman
To cite this article:Thomas W. M. Vossen, R. Kevin Wood, Alexandra M. Newman (2016) Hierarchical Benders Decomposition for Open-Pit MineBlock Sequencing. Operations Research 64(4):771-793. https://doi.org/10.1287/opre.2016.1516
Full terms and conditions of use: https://pubsonline.informs.org/Publications/Librarians-Portal/PubsOnLine-Terms-and-Conditions
This article may be used only for the purposes of research, teaching, and/or private study. Commercial useor systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisherapproval, unless otherwise noted. For more information, contact [email protected].
The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitnessfor a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, orinclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, orsupport of claims made of that product, publication, or service.
Copyright © 2016, INFORMS
Please scroll down for article—it is on subsequent pages
With 12,500 members from nearly 90 countries, INFORMS is the largest international association of operations research (O.R.)and analytics professionals and students. INFORMS provides unique networking and learning opportunities for individualprofessionals, and organizations of all types and sizes, to better understand and use O.R. and analytics tools and methods totransform strategic visions and achieve better outcomes.For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
OPERATIONS RESEARCHVol. 64, No. 4, July–August 2016, pp. 771–793ISSN 0030-364X (print) � ISSN 1526-5463 (online) http://dx.doi.org/10.1287/opre.2016.1516
© 2016 INFORMS
Hierarchical Benders Decomposition forOpen-Pit Mine Block Sequencing
Thomas W. M. VossenLeeds School of Business, University of Colorado at Boulder, Boulder, Colorado 80309, [email protected]
R. Kevin WoodOperations Research Department, Naval Postgraduate School, Monterey, California 93943, [email protected]
Alexandra M. NewmanDepartment of Mechanical Engineering, Colorado School of Mines, Golden, Colorado 80401, [email protected]
The open-pit mine block sequencing problem (OPBS) models a deposit of ore and surrounding material near the Earth’ssurface as a three-dimensional grid of blocks. A solution in discretized time identifies a profit-maximizing extraction(mining) schedule for the blocks. Our model variant, a mixed-integer program (MIP), presumes a predetermined destinationfor each extracted block, namely, processing plant or waste dump. The MIP incorporates standard constructs but also addsnot-so-standard lower bounds on resource consumption in each time period and allows fractional block extraction in a novelfashion while still enforcing pit-wall slope restrictions. A new extension of nested Benders decomposition, “hierarchical”Benders decomposition (HBD), solves the MIP’s linear-programming relaxation. HBD exploits time-aggregated variablesand can recursively decompose a model into a master problem and two subproblems rather than the usual single subproblem.A specialized branch-and-bound heuristic then produces high-quality, mixed-integer solutions. Medium-sized problems (e.g.,25,000 blocks and 20 time periods) solve to near optimality in minutes. To the best of our knowledge, these computationalresults are the best known for instances of OPBS that enforce lower bounds on resource consumption.
Keywords : industries: mining; production scheduling: deterministic; sequencing; integer programming: Bendersdecomposition; heuristic.
Subject classifications : industries: mining/metals; production/scheduling: applications; programming: integer; algorithms:Benders decomposition.
Area of review : Environment, Energy, and Sustainability.History : Received September 2014; revisions received November 2015; accepted April 2016. Published online in Articles
in Advance July 8, 2016.
1. IntroductionThe mining industry solves the open-pit mine block se-quencing problem (OPBS), primarily for strategic planningpurposes, with typical models incorporating a yearly levelof detail over a 10- to 30-year time horizon (Rojas et al.2007, Chicoisne et al. 2012, Epstein et al. 2012). This paperextends a standard integer program (IP) for OPBS to amixed-integer program (MIP) and develops a specializedsolution procedure for that MIP. Although open-pit minesmay produce diamond ore, coal, and materials other thanmetal ores, without loss of generality, we discuss OPBSin terms of mining metal ores. Figure 1 illustrates a largeopen-pit copper mine for reference.
In OPBS, a three-dimensional grid of box-shaped blocksrepresents a deposit of potentially valuable ore containingmetals such as gold or copper, along with inevitable waste.An IP or MIP for OPBS seeks a multi-period schedule forextracting (mining) and processing these blocks, a sched-ule that (i) maximizes net present value, (ii) satisfies con-straints on the shape of the mine as it evolves over time,
and (iii) satisfies constraints on resource consumption ineach time period.
Our work begins by applying lower-bounding resourceconstraints, in addition to the standard upper-bounding con-straints, to one variant of a binary (0-1) IP for OPBS(Chicoisne et al. 2012). More significantly, we relax theIP, converting it into a MIP that allows selective, frac-tional extraction of blocks: researchers typically assumethat restrictions on the shape of the mine require theuse of binary variables, but we show that our relaxedregime also satisfies those restrictions. We then developa specialized solution procedure for the new MIP that(i) defines the MIP’s linear-programming relaxation withoutexplicitly representing relaxed binary variables, (ii) solvesthat linear program using a new “hierarchical” versionof nested Benders decomposition (Ho and Manne 1974),and then (iii) incorporates that linear-programming solutionmethod within a specialized branch-and-bound heuristicthat enforces discrete relationships within the MIP throughconstraint branching. Thus, the method avoids explicit useof binary variables.
771
Vossen, Wood, and Newman: Open-Pit Mine Block Sequencing772 Operations Research 64(4), pp. 771–793, © 2016 INFORMS
Figure 1. The Bingham Canyon Mine as of 2003.
Source. http://commons.wikimedia.org/wik/File:Bingham_mine_5-10-03.jpg,accessed July 11, 2013. Notes. This copper mine is one of the largestopen-pit mines in the world. Visible in this figure are the benches, or“steps” from which ore and waste are extracted, and the haul roads,which wind down past the benches to the bottom of the pit.
As with most work on OPBS (e.g., Dagdelen andJohnson 1986, Caccetta and Hill 2003, Chicoisne et al.2012), we solve only a deterministic model, even thoughuncertainty surely plays a role in strategic mine planning(Johnson 1968). For example, price estimates for met-als 10 years in the future must have large variances, andore quality 500 meters below the Earth’s surface cannotbe known with certainty. Stochastic programming meth-ods have been suggested for open-pit mine planning (e.g.,Ramazan and Dimitrakopoulos 2007, Boland et al. 2008,Gholamnejad and Moosavi 2012), but the current state ofthe art does not permit the solution of full-scale stochas-tic programming models. Thus, we assume that (i) coresamples from the deposit (Krige 1951) and radio-imagingtechniques (Stolarczyk 1992) yield accurate deterministicestimates of each block’s weight and grade, with a block’sgrade being the percentage of metal it contains; (ii) thosevalues, together with economic forecasts, yield acceptabledeterministic estimates of the profitability of extractingeach block in each possible time period; and (iii) sourcesof uncertainty can be handled in an ad hoc manner using adeterministic model.
Several variants of OPBS exist (Espinoza et al. 2012),but our variant incorporates two key features that at leastone standard mine-planning software system also uses(Whittle 1998), namely, a “fixed cutoff grade” and “noinventorying.” Fixed cutoff grade implies that if a block’sestimated grade is at least g% for some pre-specifiedvalue g, that block is sent to a processing plant to be con-verted into salable ore; otherwise, it is sent to a wastedump. No inventorying implies that a block must be pro-cessed or dumped in the period in which it is extracted orthe block is never extracted at all.
1.1. Technical Background on OPBS
For computational reasons, OPBS normally defines extrac-tion variables of this form: xbt = 1 if block b ∈ B isextracted by (i.e., at or before) time period t ∈ T, andxbt = 0, otherwise (Johnson 1968). Strategic planners typ-ically seek an extraction schedule that covers 104–107
blocks and 10–100 time periods for a model instance thatspans 10–30 years.
If block b ∈B is not at the mine’s surface, then a uniqueblock denoted b lies directly above b; let Bb = 8b9 if bexists, and let Bb = � otherwise. Also, a specially definedset of blocks Bb may lie obliquely above b. The blocksb′ ∈ Bb ∪ Bb ≡ Bb are the direct spatial predecessorsof b, and most of the constraints in OPBS enforce spatialprecedence:
xbt − xbt ¶ 0 ∀b ∈B � Bb 6= �1 t ∈T (1)
xbt − xb′t ¶ 0 ∀b ∈B � Bb 6= �1 b′∈ Bb1 t ∈T0 (2)
That is, block b cannot be extracted by period t unless allof its direct spatial predecessors are extracted by t. Con-straints (1) and (2) are typically written as a single set ofconstraints, but we find it useful to split them into twosets because of different interpretations. In particular, con-straints (1) simply imply that a block cannot be extracteduntil the top of the block forms part of the mine’s surface,while constraints (2) mathematically enforce slope restric-tions on the pit’s walls to prevent their collapse (John-son 1968). The relationships expressed through block b’soblique predecessors b′ ∈ Bb may vary throughout thepotential mine volume depending on local characteristicsof the rock and minerals.
In addition to spatial-precedence constraints, OPBS im-plements temporal-precedence constraints, which implythat if block b is extracted by time period t < T , then thatblock must also be extracted by period t + 1:
xbt − xb1 t+1 ¶ 0 ∀b ∈B1 t = 11 0 0 0 1 T − 10 (3)
Note that both types of precedence constraints exhibit“dual network structure,” with each defining a constraint-matrix row whose nonzero coefficients are a single +1 anda single −1. Certain solution methods exploit this structurefor efficiency; see Ahuja et al. (2003) for a general dis-cussion of the topic, and see Chicoisne et al. (2012) for arecent discussion with respect to OPBS.
A solution to OPBS must also satisfy resource con-straints on production and processing in each time period(Johnson 1968). Production constraints limit the totalweight of the blocks that are extracted, while process-ing constraints limit the total weight that can be milled,i.e., crushed and refined for sale. Upper-bounding con-straints for both production and processing reflect lim-ited equipment and labor capacities, while labor contractsand requirements of exothermic processing reactions maydictate lower bounds on production and processing,respectively.
Vossen, Wood, and Newman: Open-Pit Mine Block SequencingOperations Research 64(4), pp. 771–793, © 2016 INFORMS 773
1.2. Basic Solution Approaches for OPBS
Lerchs and Grossmann (1965) describe an efficient algo-rithm for a simplified version of OPBS. Ignoring resourceconstraints and time periods in OPBS produces the ultimatepit limit problem (UPL). A solution to UPL estimates theextent of the pit beyond which no profit is possible becausefurther mining would require the extraction of excessivewaste material to reach nominally valuable ore. This modelcorresponds to a dual network with no complicating struc-ture and, therefore, solves efficiently using network-flowtechniques. In fact, the max-flow/min-cut theorem applies,and UPL may be viewed as a classical OR problem, appear-ing in numerous textbooks and research papers (e.g., Ahujaet al. 1993, pp. 721–722; Hochbaum and Chen 2000).
Johnson (1968) presents the first comprehensive descrip-tion of OPBS. He gives formulations with “at-time-t vari-ables” x′
bt (i.e., x′bt equals 1 if block b is extracted at time t,
and x′bt equals 0 otherwise), as well as formulations like
ours with “by-time-t variables” xbt . We refer the reader toLambert et al. (2014) for details on these models and ontheir direct solution by LP-based branch-and-bound meth-ods. (The abbreviation “LP” means “linear-programming”or “linear program,” depending on the context.) For ourpurposes, the key point in Lambert et al. is this: solution viabranch and bound of realistically sized OPBS models liesbeyond the capability of current-day integer-programmingsolvers. We note that Caccetta and Hill (2003) apply branchand cut to a version of OPBS and report promising resultson problems with up to 210,000 blocks and 10 time peri-ods. Reported optimality gaps are large, however, and thepaper’s lack of detail makes its results irreproducible. Thesedifficulties with direct branch-and-bound solutions, and ourdesire to avoid using heuristics and aggregation schemesthat provide no measure of solution quality (e.g., Gershon1987, Denby and Schofield 1994, Ramazan 2007, Bolandet al. 2009), motivate us to pursue a decomposition-basedsolution approach.
1.3. Decomposition Methods for OPBS
Dagdelen and Johnson (1986) appear to be the first toapply mathematical decomposition in an attempt to solveOPBS. Their Lagrangian relaxation of the model’s resourceconstraints yields a subproblem having pure dual net-work structure, which results in an integer solution tothe continuous relaxation just as the UPL model does.One must eventually find a solution that satisfies the ini-tially relaxed constraints, however, and Dagdelen and John-son’s method often fails in this regard. Other work withLagrangian relaxation and OPBS has also had limitedsuccess; for example, Akaike and Dagdelen (1999), Cai(2001), and Kawahata (2007) all have difficulty findingresource-feasible solutions.
Gleixner (2008) produces promising results using La-grangian relaxation on the variant of OPBS describedby Boland et al. (2009). This model aggregates certain
standard constructs and relaxes others, but the validityof these techniques remains unproven. Cullenbine et al.(2011) obtain high-quality solutions using a Lagrangian-based “sliding time window heuristic,” which is a type offix-and-relax heuristic (Pochet and Wolsey 2006). Lambertand Newman (2014) use Lagrangian relaxation to speedsolutions of OPBS but guarantee a solution only throughwhat may evolve into a complete OPBS model that mustbe solved by branch and bound. We seek a decompositionmethod for solving OPBS that promises to be faster thana brute-force, branch-and-bound solution of a monolithicMIP and that provides an objective measure of solutionquality.
Chicoisne et al. (2012) apply an efficient method to solvethe continuous relaxation of an OPBS IP and then apply agreedy heuristic to identify an integer solution. Because ofa strong bound from the relaxation and an effective heuris-tic, this method yields solutions with optimality gaps ofapproximately 2% on problems with up to 107 blocks and25 time periods. These authors report solution times ofa few hours on a computer having two Quad-Core IntelXeon E5420 processors. We also note that Bienstock andZuckerberg (2010) model a variant of OPBS with a vari-able cutoff grade and solve the corresponding continuousrelaxations efficiently. For instance, models with 105 blocksand 25 time periods solve in only hundreds of secondsusing a single core of a 3.2 GHz Xeon processor on a com-puter having 64 GB of memory. However, the papers men-tioned in this paragraph omit lower-bounding constraints onresource consumption, and evidence indicates that incorpo-rating both constraint types can increase computation times,even on small problems, by more than an order of magni-tude (Cullenbine et al. 2011).
We aim to take advantage of OPBS’s staircase structure,meaning that variables for period t interact directly throughconstraints only with variables for periods t − 1 and t + 1.Glassey (1971) first shows that LPs having such structurecan be solved using a nested decomposition, specifically,a nested version of Dantzig-Wolfe decomposition (Dantzigand Wolfe 1960). The key advantage of using a nesteddecomposition to solve a staircase model is that a “nestedsubproblem” involves only variables associated with a sin-gle time period. “Branch and price” extends Dantzig-Wolfedecomposition to integer problems (Johnson 1968, Barn-hart et al. 1998), but this technique seems difficult to adaptto our multistage problem. We have turned, therefore, tonested Benders decomposition (NBD), first described byHo and Manne (1974).
Benders decomposition was originally developed forsolving MIPs (Benders 1962). It views the solution of amaximizing MIP having integer variables x and continu-ous variables y as maxx∈X �4x5, where �4x5 is a piecewise-linear concave function in continuous x, and X is definedthrough a polyhedral set with integrality requirementsadded.
Vossen, Wood, and Newman: Open-Pit Mine Block Sequencing774 Operations Research 64(4), pp. 771–793, © 2016 INFORMS
The decomposition algorithm1. creates an easy-to-evaluate approximating function
¯�4x5 with ¯�4x5¾ �4x5 for all x ∈X,2. solves the “relaxed master problem” maxx∈X
¯�4x5 forx ∈X,
3. solves the LP subproblem in variables y that resultsfrom fixing x = x in the MIP,
4. extracts a dual extreme point or dual extreme ray fromthe solution to that LP to generate a new constraint calleda “Benders cut” to help refine ¯�4x5, and
5. repeats steps 2–4 until the best x found meets conver-gence criteria.
NBD extends the two-stage method to multistage LPs orto multistage MIPs with integer variables in the first stageonly. In concept, NBD views an LP in terms of a mas-ter problem and subproblem and then recursively decom-poses the subproblem into a master problem and subprob-lem using the basic ideas from standard Benders decompo-sition. We call the problem solved at any stage t of standardNBD a “nested subproblem” even though it contains con-structs of a master problem.
For simplicity, assume that each period-t nested Benderssubproblem with variables xt , t = 11 0 0 0 1 T , has a bounded,feasible solution. The following procedure then outlines astandard implementation of NBD, which solves a forwardrecursion of an LP.
1. Vector x0 defines initial conditions.2. A primal pass, in the order t = 11 0 0 0 1 T , solves a
period-t nested subproblem for primal solution xt givenxt−1. (Note that the existence of xt implies that a consistentprimal solution xt′ has been computed for t′ = 11 0 0 0 1 t − 1.)This subproblem involves only variables xt but, exceptwhen t = T , it does incorporate an approximate cost-to-go function ¯�t+14xt5, which covers the beginning of periodt + 1 through the end of period T .
3. A dual pass, in the order t = T 1 0 0 0 12, solves aperiod-t nested subproblem for dual solution Ït to generatea new Benders cut that refines the cost-to-go function forthe nested subproblem in period t − 1. (Actually, the solu-tion to the period-T nested subproblem in the primal passyields the initial dual-pass result ÏT .)
4. Steps 2 and 3 are repeated until the pessimistic boundfrom step 2 and the optimistic bound from step 3 aresufficiently close.(Note that some authors use “forward recursion” and “back-ward recursion” to mean what we call “primal pass” and“dual pass,” respectively.)
It is possible to reorganize computations above, forinstance, by iterating between a primal solution for period tand a dual solution for period t + 1 until some local con-vergence criterion is reached, then iterating between t + 1and t + 2, etc. However, the outline above describes thesubproblem-processing method that both Wittrock (1985)and Gassman (1990) find most efficient for implementingNBD and that we therefore adopt or adapt, as needed. Weuse Wittrock’s term for this method, “FASTPASS.”
Staircase structure lends nested decomposition its com-putational advantage, but this structure is also its Achillesheel. If strong linkages exist between distant time periods,eventually those linkages must be represented by generat-ing and applying Benders cuts in many time periods. Slowconvergence results. We seek to improve the empirical con-vergence rate using several techniques.
First, we formulate the model using cumulative vari-ables, that is, variables that are cumulative over time. Ofcourse, cumulative extraction variables in OPBS formula-tions are actually standard: xbt = 1 if block b is extractedby time period t, and xbt = 0, otherwise.
Next, we add redundant, aggregate resource constraintsto help guide the decomposition. For example, the decom-position procedure’s first subproblem might aggregate alltime periods into a single cumulative period and, inessence, ask, “What is the optimal ‘relaxed open-pit mine’that could be excavated in a single period if each resourceconstraint aggregates resource availability over the entiretime horizon?” Intuitively, this provides immediate, globalinformation to subsequent subproblems. By contrast, thefirst primal pass of a standard forward recursion wouldgreedily excavate the “relaxed mine” from one period to thenext, with global information appearing only slowly as theprocedure refines approximating cost-to-go functions overmany iterations. (Section 3.6 covers this topic further andcompares our techniques to others in the literature.)
Finally, we show that the standard recursion used to cre-ate a multistage Benders decomposition is a special caseof a tree decomposition: a standard decomposition recur-sively decomposes a multistage problem into a master prob-lem and a subproblem, while the tree decomposition canrecursively decompose that problem into a master prob-lem and two subproblems. This more general decomposi-tion framework may be viewed in terms of a binary treeand, consequently, resembles the decomposition schemeproposed by Kallio and Porteus (1977) for solving a setof linear equations having a tree structure. Kallio andPorteus intend for their decomposition to improve com-putational efficiency but provide no supporting, computa-tional results. We also note that these authors assume agiven tree structure, whereas we create various tree struc-tures through problem reformulations. In the literature ondecomposition for optimization, only the work by Entriken(1989) seems closely related to ours. Entriken proposes aframework for decomposing an LP that could, in principle,yield a formulation having a general tree structure. He pro-vides computational results only for sequentially structureddecompositions, however; see also Entriken (1996).
For simplicity, we use the phrase hierarchical Ben-ders decomposition (HBD) to refer to the combination ofall three techniques just described, i.e., cumulative vari-ables, aggregate constraints, and a tree-structured decom-position. While somewhat specialized, HBD should alsoapply to a number of multistage production-schedulingproblems in the literature, including production-planning
Vossen, Wood, and Newman: Open-Pit Mine Block SequencingOperations Research 64(4), pp. 771–793, © 2016 INFORMS 775
problems (Gabbay 1979), production-distribution problems(Brown et al. 2001), short-term scheduling for forest har-vesting (Karlsson et al. 2003), and short-term open-pit minescheduling (Eivazy and Askari-Nasab 2012). The key isthat production efficiencies or yields should not changesubstantially over time.
Eventually we need a solution to a MIP, not just an LP.For the same reason that Benders decomposition appliesto problems with integer variables in the first stage only,NBD and HBD apply (directly) to staircase MIPs with inte-ger variables in a single stage only (e.g., Wollmer 1980).To address this limitation, we develop a branch-and-boundheuristic that produces high-quality binary solutions for avariety of test problems. For efficiency, the methods repre-sent binary variables only implicitly.
1.4. Outline of the Remainder of the Paper
Section 2 begins the remainder of this paper by describ-ing a new MIP for OPBS and justifying the changes frommore standard models. Section 3 demonstrates how stan-dard NBD applies to solve the LP relaxation of the MIP andthen generalizes that method to HBD; Section 4 presentscorresponding computational results. Based on solving LPrelaxations with HBD, Section 5 devises a branch-and-bound heuristic to obtain mixed-integer solutions to OPBS;Section 6 presents corresponding computational results.Section 7 summarizes all computational results, and Sec-tion 8 concludes the paper.
2. A New Model for OPBSBeginning with a standard IP for OPBS, this section devel-ops a new MIP for that problem; defines a useful restrictionand a useful relaxation of the MIP; and then validates thekey, novel feature of the MIP, which is the modeling offractional block extraction.
2.1. A Mixed-Integer Programming Formulation
We use the “C-PIT” IP of Chicoisne et al. (2012) as a start-ing point and describe a new MIP for OPBS by (i) imposinglower as well as upper bounds on resource consumptionin each period and (ii) allowing selective, fractional blockextraction. Normally, spatial-precedence constraints (2),together with strict integrality of variables, enforce pit-wall slope restrictions, but we show that these restrictionsremain enforced even with the relaxation implied by (ii).As is standard, our model represents the potential mine vol-ume as a three-dimensional grid of blocks having commondimensions, with blocks stacked directly on top of eachother. We denote the new model simply as “MIP.”
Indices and Index Sets
b ∈B extractable blocksBb ⊂B all direct spatial predecessors of block bb ∈Bb if it exists, the direct spatial predecessor that
lies above block b
Bb ⊂Bb Bb = 8b9 if block b exists, and Bb = �,otherwise
Bb ⊂Bb Bb\Bb, i.e., the oblique direct spatialpredecessors of block b
t ∈T time periods defining the time horizon;T= 811 0 0 0 1 T 9
r ∈R production and processing resources
Data: [units]
v′bt net present value of block b if extracted in
period t [dollars]vbt v′
bt − v′b1 t+1, with v′
b1T+1 ≡ 0 for all b ∈Bqrb consumption of resource r associated with the
extraction of block b [tons] (Note that qrb = 0 ifr corresponds to processing a waste block b.)
qLrt 4q
Urt5 minimum (maximum) consumption limits for
resource r in time period t [tons]
Variables: [units, if defined]
xbt 1 if block b is completely extracted by time period t,0 otherwise
ybt fraction of block b extracted by time t; nominally,yb0 ≡ 0 for all b ∈B
Formulation:
MIP2
�∗
MIP = maxx1y
∑
b∈B
∑
t∈T
vbtybt (4)
s0t0 −∑
b∈B
qrb4ybt−yb1t−15¶−qLrt
∀r ∈R1 t∈T (5)∑
b∈B
qrb4ybt−yb1t−15¶qUrt ∀r ∈R1 t∈T (6)
−4ybt−yb1t−15¶0 ∀b∈B1 t∈T (7)
ybt−xbt¶0 ∀b∈B �Bb 6=�1 t∈T (8)
ybt−yb′t¶0
∀b∈B �Bb 6=�1 b′∈Bb1 t∈T (9)
xbt−ybt¶0 ∀b∈B1 t∈T (10)
xbt ∈80119 ∀b∈B1 t∈T (11)
ybt¾0 ∀b∈B1 t∈T (12)
yb0 ≡0 ∀b∈B (13)
Through its objective function (4), MIP seeks to maxi-mize the total net present value of extracted blocks. Foreach time period, constraints (5) and (6) restrict minimumand maximum resource consumption, respectively. Con-straints (7) enforce (relaxed) temporal-precedence relation-ships for each block: if a fraction of block b is extractedby time t − 1, then at least that fraction must be extractedfrom block b by time t.
Vossen, Wood, and Newman: Open-Pit Mine Block Sequencing776 Operations Research 64(4), pp. 771–793, © 2016 INFORMS
Note that some blocks at a mine’s surface may be “par-tial” at the beginning of time period 1 because the undis-turbed surface is uneven or because some partial extractionhas already taken place. Although we assume that yb0 = 0for all computational tests, partial blocks could be handledby fixing certain instances of yb0 to appropriate nonzerovalues in (13).
In effect, a standard OPBS model enforces spatial prece-dence through constraints (8) and (9) while using onlybinary variables: block b cannot be extracted until allblocks b′ ∈ Bb ∪ Bb have been extracted completely. OurOPBS model MIP uses a combination of binary and con-tinuous variables to enforce the following, relaxed, spatial-precedence requirements:
• assuming block b lies above block b, constraints (8)imply that block b cannot begin to be extracted until b iscompletely extracted; and
• assuming block b has some oblique spatial predeces-sors, i.e., Bb 6= �, constraints (9) imply that the fractionof block b that is extracted by time t cannot exceed thefraction that is extracted by time t from any b′ ∈ Bb.
We validate this use of fractional block extraction in Sec-tion 2.3 after describing important ways that we restrict andrelax MIP.
2.2. Restricting and Relaxing MIP
Later, we need to compare solutions of MIP to those ofa standard IP for OPBS. Since we derived MIP fromsuch an IP, it is easy to recreate it: (i) restrict all vari-ables ybt to be binary; (ii) replace all xbt with ybt; (iii) deleteconstraints (10)–(12), which have become redundant; and(iv) call the resulting model “IP.”
As is standard, we begin the solution process for MIP byfirst solving its LP relaxation. Actually, we solve a specialform of this relaxation, denoted RMIP: this is identical tothe LP relaxation of IP, just described
After solving RMIP, we apply a branch-and-boundheuristic, which dynamically and implicitly enforces (8)and (10) in MIP, ensuring that these restrictions holdfor every block b such that Bb 6= �. Specifically, ybt >0 ⇒ ybt = 1, whenever b exists. The resulting solution4y∗
11 0 0 0 1 y∗
T 5 is said to be “MIP valid” for MIP.
2.3. Validating Fractional Block Extraction
The validity of constraints (8) as relaxed versions of con-straints (1) is clear: no fraction of a block b can be extracteduntil the block b, which lies directly above b, is com-pletely extracted. That statement is true whether the frac-tion in question lies between 0 and 1 or the “fraction”must be exactly 0 or 1. The validity of constraints (9) asrelaxed versions of (2) is less clear, however. This sec-tion demonstrates the geometrical validity of the fractionalblock extraction modeled in MIP and then solves somesmall instances of MIP and IP to investigate practicalimplications. We note that Gershon (1983) also considers
fractional block extraction but uses a more restrictive def-inition: fractional extraction of a block b is allowed pro-vided that all blocks b′ ∈Bb have been extracted fully.
2.3.1. Theory. In IP, each spatial-precedence con-straint defined by (2) restricts the local pit-wall slope angle.We show here that each constraint defined by (9) restrictsthe local slope in a solution to MIP to an angle that is nosteeper than that enforced by the corresponding constraintin IP; Figure 2 illustrates. We demonstrate informally, first,assuming that each block is a cube with each side havinga length of one in arbitrary units.
Figure 2(a) reflects a standard set of spatial-precedencerelations: block b0 cannot be extracted until each blockb′ ∈ Bb0
∪Bb0= 8b09∪ 8b11 b21 b31 b49 is extracted. We sim-
plify to the two-dimensional model of Figure 2(b) so thatblock b0 has oblique predecessors Bb0
= 8b11 b29, only.Dropping the subscript t for simplicity, these constraints
define the standard spatial-precedence relationships for thetwo-dimensional example:
xb0− xb0
¶ 0 (14)
xb0− xb1
¶ 0 (15)
xb0− xb2
¶ 00 (16)
Assuming that the block below b0 is not extracted, standardslope restrictions associated with b0 may be interpreted asfollows (see Figure 2(c)):
(i) constraint (14) requires that b0 be extracted com-pletely before b0 is extracted;
(ii) constraint (15) requires that the slope �01, measuredfrom the center of the extracted face of b0 to the centerof the extracted face of b1, not exceed arctan4d01/D015 =
arctan41/15= 45�; and, similarly,(iii) constraint (16) requires that the slope �02, which
is analogous to �01, not exceed arctan4d02/D025 =
arctan41/15= 45�.Thus, constraints (14)–(16) enforce pit-slope restrictions
of “at most 45�.”Now, when allowing fractional block extraction in
MIP (see Figure 2(d)), the following constraints replace(14)–(16), respectively:
yb0− xb0
¶ 0 (17)
yb0− yb1
¶ 0 (18)
yb0− yb2
¶ 00 (19)
If yb0= 0, then block b0 has not been extracted, and these
constraints impose no restrictions on the extraction of b0,b1, or b2. (An analogous situation arises in the all-binarymodel when xb0
= 0; see (14)–(16).) If yb0= 1, and the
block below b0 has not been extracted, then standard sloperestrictions are enforced; i.e., yb1
= xb0= yb2
= 1. Thus, weonly need to ensure that, when 0 < yb0
< 1,(i) block b0 is completely extracted;
Vossen, Wood, and Newman: Open-Pit Mine Block SequencingOperations Research 64(4), pp. 771–793, © 2016 INFORMS 777
Figure 2. Fractional extraction maintains slope restrictions.
(a)
(d)(c)
(b)
b1 b0 b2
b0 d02d01
D02D01
�01 = 45.0º �02 = 45.0º
b2
b0
b4
b3
b1
b0
b1 b0 b2
b0
b1 b0 b2
b0d01 d02
D02D01
�01 = 45º �02 = 36.9º
Notes. Each block side has unit length. (a) Shows the six blocks, b0, b0, b11 0 0 0 1 b4, that are involved in maintaining slopes as measured from b0. (b)Simplifies to two dimensions for purposes of illustration. (c) Shows how slope angles are defined from the blocks’ extracted faces (i.e., the bottom faces ofthe blocks) when fractional extraction is disallowed. Assuming the blocks (not shown) directly beneath b0, b1, and b2 have not been extracted, �01 measuresthe slope from b0 toward b1 and �02 measured from b0 toward b2. The illustrated angles of 45� are the maximum allowable: �01 = arctan4d01/D015 =
arctan41/15= 45� and �02 = arctan4d02/D025= arctan41/15= 45�. (d) Shows slope angles �01
and �02
, corresponding to �01 and �02, respectively, but withfractional extraction allowed; unshaded regions have been extracted, while shaded regions have not. Now we see that �
01= arctan4d01/D015= arctan41/15=
45� = �01, but �02
= arctan4d02/D025= arctan40075/15= 3609� < 45� = �02.
(ii) the slope �01
between extracted faces of b0 and b1
does not exceed 45�; and(iii) the analogous slope �
02does not exceed 45�.
Next, Figure 2(d) illustrates the case in which yb0= yb1
=
0051 yb0= 100, and yb2
= 0075. Now,(i) is satisfied in general through constraint (17);(ii) is satisfied because
�01 = arctan4d01/D015= arctan41/15= 45� = �01; and(iii) is satisfied because
�02
= arctan4d02/D025= arctan40075/15= 3609� < 45�.In general, with respect to (ii) and (iii) when 0 < yb0
< 1,a solution to MIP defines a slope � between b0 andany block b′ ∈ Bb such that � = arctan441 + 41 − yb′5 −
41 − yb055/15¶ 45� because (9) ⇒ 1 − yb′ + yb0
¶ 1. Thus,
the relaxed model also enforces pit-slope restrictions of “atmost 45�.”
In the following theorem, we extend the discussion aboveto more general slope relationships and more general blockgeometries.
Theorem 1. In allowing fractional block extraction, an in-stance of MIP ensures that pit-wall slope angles do notexceed those that the corresponding (all-binary) instanceof IP would enforce, provided that each block has commondimensions and a rectangular base.
Proof. Given the discussion above, it suffices to show thatfor any pair of blocks 4b1 b′5 such that b′ ∈ Bb, the relevantconstraint from (9) enforces an angle between the extractedfaces of b and b′ that does not exceed the angle enforced
Vossen, Wood, and Newman: Open-Pit Mine Block Sequencing778 Operations Research 64(4), pp. 771–793, © 2016 INFORMS
Figure 3. Figure for proof of Theorem 1.
h
h
d = k ·h
D
dBlock b
90º
Block b�∈�b
h(1 – yb�t)
h(1 – ybt)
�
�
Notes. Block shapes are exaggerated for clarity; h denotes the blocks’common height. The shaded portions of the blocks illustrate unextractedportions that would be valid in MIP. Assuming an all-integer version ofMIP, � indicates the enforced angle from the extracted face of b to theextracted face of b′ ∈ Bb . For MIP, � indicates the corresponding anglewhen the extracted fraction of block b is ybt and the extracted fractionof b′ is yb′ t . Note that ybt < yb′ t in the figure, satisfying constraints (9) inMIP.
constraint in (2). Figure 3 illustrates a general case in which(i) each block has height h; (ii) the bottom center of blockb is located at general grid coordinates (x, y, z); and (iii) thebottom center of block b′ is located at coordinates 4x +
Dx1y +Dy1 z + d5 such that Dx ¾ 0, Dy ¾ 0, Dx +Dy > 0,and d = k ·h for some positive integer k. (None of x, y, z,Dx, or Dy is indicated in the figure.)
Now, D = 4D2x + D2
y51/2 > 0 defines the horizontal dis-
tance between the blocks’ centers, as indicated in the figure,and we know from earlier discussion that constraints (2)in the all-integer version of MIP enforce a slope of � =
arctan4d/D5, when that slope is defined. For MIP, let thecorresponding angle be �, and assume that this angle isdefined in period t. Thus,
� = arctan4d/D5 (20)
= arctan44d+ 41 − yb′t5h− 41 − ybt5h5/D5 (21)
= arctan44d+ 4ybt − yb′t5h5/D5 (22)
¶ arctan4d/D5 because constraints (9)require that ybt − yb′t ¶ 0 (23)
= �0 � (24)
2.3.2. Practical Implications. Here, using data setsfor five different open-pit mine scenarios, we compare solu-tions of MIP to solutions of IP. The data sets can createproblem instances that cover 10,819, 18,300 and 25,620blocks, for 1 to 20 time periods, although we use a max-imum of 10 time periods here. Cullenbine et al. (2011)use the same data sets, plus one denoted “BD10819F,”
and we follow that paper’s naming conventions, with eachdata set’s label specifying the relevant number of blocks.(The current section omits results for BD10819F simplybecause of space limitations in Table 1. Subsequent compu-tational results for decomposition-based methods do coverthat data set, however.)
We consider only small values of T so that the problemscan be solved using LP-based branch and bound, that is,without requiring the decomposition techniques developedlater in the paper. Computations are performed on a LenovoW541 laptop computer having a 64-bit, quad-core, Intelprocessor running at 2.9 GHz. The computer has 16 GBRAM and runs the Windows 7 Professional operating sys-tem. A C++ program generates all models and CPLEX12.6 (IBM Corp. 2014) solves them. We override CPLEX’sdefault parameters in five different ways:
(i) the solver may use at most four threads (Threads = 4);(ii) a relative optimality tolerance of 0.1% applies
(EpGap = 0.001);(iii) computations are limited to 7,200 seconds of
elapsed time (TiLim = 7,200);(iv) the barrier algorithm solves root-node LPs
(RootAlg invokes BarAlg); and(v) because of the cumulative nature of the models’
variables, branching priorities for xbt are set to t, i.e.,priority increases with t.
Table 1 displays results and “Notes” provide detailedexplanations of the table’s entries. We highlight the follow-ing points from these results:
• Average profit for a solution to MIP compared to IPimproves by at least 1.0% but no more than 1.9%; theinability to solve many instances of IP accurately makesmore precise statements impossible.
• No clear trend in improved profit for MIP versus IPappears as T increases, i.e., as the mine pit expands.
• Because counting all fractional variables ybt in MIPwould imply some double counting when 0 < ybt ≈
yb1 t+1 < 1, the table lists the number of fractional variablesonly for the last time period, T . No clear trend in that num-ber appears as T increases.
• The number of fractional variables ybT in MIP mayconstitute a small percentage of the total number of posi-tive variables in period T (less than 3% for the two largestinstances of BD25620A), or it may constitute a substan-tial percentage (almost 50% for the smallest instance ofBD18300A).
• Despite having more variables and constraints, theflexibility provided by fractional block extraction in MIPmakes that model much easier to solve than IP.
2.3.3. Conclusion. The qualities of solutions to MIPdeserve further investigation, but Section 2.3 has shownthat (i) MIP’s fractional block extraction leads to extractionschedules that satisfy pit-wall slope restrictions; (ii) solu-tions to MIP may yield profits that are 1%–2% higher thanwith IP; and (iii) MIP has computational advantages over
Vossen, Wood, and Newman: Open-Pit Mine Block SequencingOperations Research 64(4), pp. 771–793, © 2016 INFORMS 779
Table1.
Val
idat
ing
frac
tiona
lbl
ock
extr
actio
n:C
ompa
ring
solu
tions
ofMIP
andIP
prod
uced
bya
stan
dard
bran
ch-a
nd-b
ound
met
hod.
MIP
IPO
bs.
Min
.So
ln.
B&
BA
vg.
Soln
.B
&B
Opt
.pr
ofit
profi
tM
odel
Var
s.b
Con
s.c
timed
node
se�
f MIP
One
sgFr
acs.
hfr
acs.
itim
edno
dese
�j IP
�k IP
One
sgga
plin
cr.m
incr
.n
nam
eaT
(num
.)(n
um.)
(sec
.)(n
um.)
($)
(num
.)(n
um.)
(num
.)(s
ec.)
(num
.)($
)($
)(n
um.)
(%)
(%)
(%)
BD
1081
9A2
4312
8512
9193
741
040
5159
3164
937
011
15505
2904
251
5146
4113
551
4681
601
398
001
203
203
486
1569
2701
693
2200
80
8131
2144
277
392
2300
7120
103
7781
769
8127
0156
181
3151
961
775
005
005
000
612
9185
341
1144
954
109
091
4271
565
1108
217
72905
7120
205
9914
7491
3251
300
9138
9141
911
141
007
101
004
817
3113
755
2120
521
4510
831
927
1012
5217
6011
463
219
2704
7120
303
7716
0310
117819
8910
1248
1853
1149
700
700
700
010
2161
421
6921
961
2149
003
010
192110
6711
783
150
1500
7120
209
1110
6010
1808
1960
1019
0113
1411
839
008
100
002
BD
1830
0A1
3616
0590
1536
500
381912
2716
0237
333300
300
71818
0119
8218
1811
1748
3800
120
320
22
7312
0919
9137
231
0872
3714
4813
4744
341700
3142
902
3119
4336
163815
9936
1638
1599
7600
120
220
23
1091
813
3081
208
9901
137
5413
4213
9281
531707
7120
603
4187
852
1734
1812
5410
7414
6611
620
530
000
54
1461
417
4171
044
4130
811
342
7010
3311
7414
018
405
7120
108
1156
968
1242
1623
6917
7317
8915
520
220
600
45
1831
021
5251
880
3149
102
7139
083
1970
1633
181
2440
871
2020
321
949
8017
9111
6583
1675
1968
193
306
309
004
BD
1830
0B1
3616
0590
1536
102
022
197110
2837
220
010
40
2216
7212
8922
1677
1547
3800
010
310
32
7312
0919
9137
252
050
4016
2211
4875
310
511
4040
711
033
4012
3217
8840
1268
1474
7600
110
000
93
1091
813
3081
208
3906
05511
7016
0111
17
203
4198
709
2114
254
162612
3954
1661
1314
115
001
100
009
414
6141
741
7104
410
408
06716
6413
4814
716
400
7120
106
1180
366
1966
1310
6712
7618
0815
300
510
000
65
1831
021
5251
880
4760
70
7814
7519
0418
57
104
7120
203
623
7712
9416
6578
112817
4219
310
110
500
4B
D25
620A
15112
4513
3197
220
60
2810
8117
5735
440
020
80
2716
7211
6927
1683
1547
3800
010
510
42
1021
489
2931
564
8006
04915
5911
0174
420
034
901
732
4818
8013
0548
1928
1136
7700
110
410
33
1531
733
4531
156
3270
10
6715
7718
9811
07
203
7120
200
4180
766
1660
1574
6617
7816
4411
500
210
410
24
2041
977
6121
748
2970
00
8310
7011
8015
04
100
7120
206
2141
281
193710
8382
1379
1219
154
005
104
008
525
6122
177
2134
011
7660
30
9612
9113
3618
75
100
7120
300
2137
295
1052
1790
9518
3710
8119
200
810
300
5B
D25
620B
15112
4513
3197
260
50
1219
5311
9035
440
070
40
1215
9313
4012
1594
1550
3800
020
920
92
1021
489
2931
564
7105
02316
6017
6451
301500
7120
203
4810
4323
1021
1102
2313
0218
7876
102
208
105
315
3173
345
3115
620
103
03315
1918
0992
3010
0071
2050
336
1627
3216
0418
2833
1132
1682
114
106
208
102
420
4197
761
2174
811
7910
139
342
1592
1824
109
681700
7120
706
2011
6041
1284
1258
4211
1312
4415
320
030
110
15
2561
221
7721
340
1162
806
156
5018
7816
1217
681
1602
7120
302
5175
749
1191
1860
5014
7710
7519
120
630
400
8
aN
ame
ofth
em
odel
,w
hich
indi
cate
sth
enu
mbe
rof
bloc
ksth
atit
cont
ains
.bN
umbe
rof
vari
able
sin
MIP
;IP
has
50%
few
er.
c Num
ber
ofco
nstr
aint
sin
MIP
;IP
aver
ages
17%
few
er.
dW
all-
cloc
ktim
efo
rC
PLE
Xto
solv
eth
em
odel
inst
ance
,MIP
orIP
,but
limite
dto
7,20
0se
cond
s.T
his
time
omits
abou
tth
ree
seco
nds
ofov
erhe
adfr
omm
odel
gene
ratio
n.e N
umbe
rof
node
sin
CPL
EX
’sbr
anch
-and
-bou
ndtr
ee.
f Bes
tso
lutio
n’s
obje
ctiv
eva
lue
forMIP
.B
ecau
seal
lso
lutio
nsto
MIP
satis
fyth
etig
htop
timal
ityto
lera
nce
of0.
1%,
noco
lum
nfo
r“o
bser
ved
rela
tive
optim
ality
gap”
ispr
ovid
edan
dco
mpu
tatio
nsof
incr
ease
dpr
ofit
forMIP
vers
usIP
assu
me
that
�MIP
=�
∗ MIP
.gN
umbe
rof
bina
ryva
riab
lesxbT
such
that
xbT
=10
hN
umbe
rof
frac
tiona
lva
riab
les,
i.e.,
num
ber
ofco
ntin
uous
vari
able
sy b
Tsu
chth
at0<y b
T<
10i N
umbe
rof
frac
tiona
lva
riab
lesy b
T.
j Bes
tso
lutio
n’s
obje
ctiv
eva
lue
forIP
,al
thou
ghth
ism
ayno
tsa
tisfy
the
desi
red
rela
tive
optim
ality
tole
ranc
eof
001%
.kB
est
uppe
rbo
und
on�
∗ IP.
l Rel
ativ
eop
timal
ityga
pob
serv
edin
solu
tion
toIP
.m
Obs
erve
dpe
rcen
tage
incr
ease
inpr
ofit
forMIP
vers
usIP
:10
0%×4�
MIP
−�IP5/�IP
.nM
inim
umpo
ssib
lepe
rcen
tage
incr
ease
inpr
ofit
forMIP
vers
usIP
,w
hich
hypo
thes
izes
that
aso
lutio
nto
IPca
nbe
foun
dw
ithob
ject
ive
valu
e�IP
:10
0%×4�
MIP
−�IP5/�IP
.
Vossen, Wood, and Newman: Open-Pit Mine Block Sequencing780 Operations Research 64(4), pp. 771–793, © 2016 INFORMS
IP, at least when trying to solve those models by LP-basedbranch and bound. As discussed later, we have attemptedto solve IP using the decomposition-based heuristic thatsuccessfully solves MIP. The branch-and-bound portion ofthat heuristic requires orders of magnitude more time whenoperating on IP than when operating on MIP, however.Thus, MIP also appears to have computational advantagesover IP in the context of decomposition.
3. Solving RMIP by DecompositionThis section describes how to solve RMIP using standardnested Benders decomposition and then develops our newhierarchical variant of that decomposition, HBD. A novelapplication of aggregate resource constraints, made possi-ble by the use of cumulative variables, turns out to be cru-cial for the computational effectiveness of HBD. For easeof exposition, we add a dummy time period t = T + 1 andpresent RMIP using matrix notation.
Additional Notation
T+ 811 0 0 0 1 T + 19T0+ 801 0 0 0 1 T + 19T T + 1
y01 yT initial conditions and upper bound on ending condi-tions for the mine, respectively; y0 = 0 and yT = 1may be assumed
yt 4y1t1 0 0 0 1 y�B�t5> for all t ∈ T0+; however, y0 ≡ y0
and yT ≡ yTvt 4v1t1 0 0 0 1 v�B�t5 for all t ∈T0+; however, v0 = vT = 0qt 44−qL
1t1 0 0 0 1−qL�R�t51 4q
U1t1 0 0 0 1 q
U�R�t55
> for all t ∈T+,where −qL
b1 Tand qU
b1 Tare large enough to make
constraints (26) vacuous when t = T
The continuous relaxation of MIP then has the followingform:
RMIP2 �∗
RMIP = maxy010001yT
∑
t∈T+
vtyt (25)
s0t0 A4yt−yt−15¶qt ∀t∈T+ (26)
−I4yt−yt−15¶0 ∀t∈T+ (27)
Htyt¶0 ∀t∈T (28)
yt¾0 ∀t∈T (29)
y0 ≡ y0 (30)
yT ≡ yT 1 (31)
where1. the matrix A derives from constraints (5)–(6);2. the �B� × �B� identity matrix I in (27) derives from
constraints (7);3. the matrix Ht derives from constraints (9) as well as
the constraints that result from aggregating pairs of con-straints taken from (8) and (10); and
4. we define a feasible solution y = 4y01 0 0 0 1 yT 5 toRMIP to be MIP-valid, if there exists x = 4x01 0 0 0 1 xT 5such that 4x1 y5 is a feasible solution to MIP.
Note that upper bounds yt ¶ 1 for all t are implied by (27)and (31).
Observe that Ht is actually stationary in our applica-tion (i.e., Ht = H for all t), but it need not be. Thematrix A need not be stationary for standard NBD, butwe exploit stationarity of A when using aggregate resourceconstraints. (The final paragraph of Section 3.6 discussesextensions to “nearly stationary” matrices At .) To simplifylater descriptions of NBD and HBD, we develop neces-sary notation here in the context of standard, “non-nested”Benders decomposition applied to a staircase LP.
To begin, let t1 t ∈ T0+, t < t, and suppose that y t andyt are given such that (i) y0 ¶ y t ¶ yt ¶ yT , (ii) Ht y t ¶ 0,and (iii) Ht yt ¶ 0. That is, except for not necessarily beingMIP-valid, y t defines a valid pit through time period t,which is “nested” inside of the pit defined by yt .
Given the above conditions, the following model gener-alizes the standard concept of a cost-to-go function to acost-to-operate function, which covers the end of period tto the beginning of period t:
�∗4y t1 yt5
≡ maxyt 10001yt
t−1∑
t=t+1
vtyt (32)
s.t. 42651 4275 for t = t + 11 0 0 0 1 t, only (33)
42851 4295 for t = t + 11 0 0 0 1 t − 1, only (34)
yt ≡ y t (35)
yt ≡ yt0 (36)
Remark 1. Strictly speaking, �∗4y t1 yt5 should display anadditional identifier such as a subscript 6 t1 t7, but the rel-evant information will be clear from the function’s argu-ments so we omit it. Note also that we have dropped thesubscript RMIP for notational simplicity.
Remark 2. The actual implementation of RMIP uses elas-tic resource constraints, so given conditions (i)–(iii) justspecified, �∗4y t1 yt5 always has a finite value. For simplicityhere, we omit explicit representation of elastic constructs,but Appendices A and B do cover the details.
Remark 3. The optimal objective value for RMIP is�∗RMIP = �∗4y01 yT 5.
Now, for any t1 t′1 t ∈T0+ with t < t′ < t, define
Yt′4y t1yt5
=
y t¶yt′ ¶ yt
∣
∣
∣
∣
∣
∣
∣
Ht′yt′ ¶01
and A4yt′ − yt′−15¶qt′ if t′ = t+1
and A4yt′+1 −yt′5¶qt′+1 if t′ = t−1
0
(37)
Note that bounds y t ¶ yt′ ¶ yt must hold for any modelusing cumulative variables yt′ ; see constraints (9).
Vossen, Wood, and Newman: Open-Pit Mine Block SequencingOperations Research 64(4), pp. 771–793, © 2016 INFORMS 781
Next, consider the optimal solution of RMIP computedusing two functions of yt′ :
�∗4y01yT 5 = �t′4y01yT 5 (38)
≡ maxyt′ ∈Yt′ 4y01 yT 5
vt′yt′ +�∗4y01yt′5+�∗4yt′ 1yT 50 (39)
We know that constraint (37) is valid because yt′ mustsatisfy the constraints of a (relaxed) pit that lies “nestedbetween” the pits defined by y0 and yT . From standardLP theory, we also know that the functions �∗4y01yt′5and �∗4yt′ 1 yT 5 are piecewise linear and concave for yt′ ∈
Yt′4yt01 yT 5.To solve RMIP, standard Benders decomposition would1. view �∗4y11yt′5 + �∗4yt′ 1 yT 5 as a single, piecewise-
linear, concave function, say �∗4y01yt′ 1 yT 5;2. replace �∗4y01yt′ 1yT 5 with a piecewise-linear approx-
imating function, say ¯�4y01yt′ 1yT 5, such that ¯�4y01yt′ 1yT 5¾�∗4y01yt′ 1yT 5 for all yt′ ∈Yt′ ; and
3. solve a sequence of nested subproblems that succes-sively improves the approximating function and convergesto an optimal solution. (Of course, the decomposition algo-rithm typically terminates when the best solution found sat-isfies a prespecified optimality criterion.)
Maintaining separate approximating functions for�∗4y01yt′5 and �∗4yt′ 1 yT 5 leads to a multicut master prob-lem defined with respect to two separate subproblems(Birge and Louveaux 1988). Two special cases arise, how-ever: if t′ = 1, then (39) simplifies to
�14y01 yT 5= maxy1∈Y14y01 yT 5
v1y1 + �∗4y11 yT 51 (40)
and if t′ = T , then (39) simplifies to
�T 4y01 yT 5= maxyT ∈YT 4y01 yT 5
vT yT + �∗4y01yT 50 (41)
Section 3.5 shows how to recursively decompose bothfunctions �∗4y01yt′5 and �∗4yt′ 1 yT 5 to create a generaltree decomposition. First, however, Section 3.1 presents astandard forward recursion for NBD, Section 3.2 indicateshow applying aggregate resource constraints may improvethe decomposition algorithm’s efficiency, and Sections 3.3and 3.4 describe simple variants of NBD that make use ofthose constraints. From this point on, “subproblem” implies“nested subproblem.”
3.1. A Forward Recursion for Nested BendersDecomposition (FBD)
RMIP exhibits a staircase structure, which seems idealfor solution through a nested decomposition, either nestedDantzig-Wolfe decomposition (Glassey 1973) or NBD (Hoand Manne 1974). We implement NBD because construct-ing a MIP valid solution for MIP from the continuoussolution to RMIP seems easier with NBD. For reference,then, this section describes a standard version of NBD. We
note that NBD was first proposed for solving deterministicproblems but has become particularly important for solv-ing multistage stochastic programs (Birge 1997). Perhapsthis fact will make NBD useful for solving certain stochas-tic versions of OPBS, for example, with probabilisticallymodeled net present values for blocks.
The following equations describe a forward recursion ofRMIP, which leads to a “forward NBD” (FBD). This is theclassical nested Benders decomposition of Wittrock (1985).
�∗4y01yT 5
= maxy1∈Y14y01 yT 5
v1y1 +����*0
�∗4y01y15+�∗4y11yT 5 (42)
= maxy1∈Y14y01 yT 5
v1y1 +�24y11yT 5 (43)
= maxy1∈Y14y01 yT 5
v1y1 +
{
maxy2∈Y24y11 yT 5
v2y2 +�34y21yT 5}
(44)
000
= maxy1∈Y14y01 yT 5
v1y1 +
{
maxy2∈Y24y11 yT 5
v2y2
+
{
···
{
maxyT ∈YT 4yT−11 yT 5
vT yT +����*
0�∗4yT 1yT 5
}
···
}}
(45)
More simply, the FBD recursion may be summarizedthrough the repeated application of the following relation-ships, starting with t = 0 and with fixed values t = T , y0 =
y0, and yT = yT :
�∗4yt1yt5 = �t+14yt1 yT 5 (46)
= maxyt+1∈Yt+14yt 1 yT 5
vt+1yt+1 + �∗4yt+11 yT 50 (47)
To exploit the FBD recursion, an FBD algorithm re-places each function �t+14yt1 yT 5 with an upper-bound-ing, piecewise-linear, concave approximation ¯�k
t+14yt1 yT 5,which it improves in each iteration k using standard opti-mality and feasibility cuts (although our implementationrequires only optimality cuts). In our FASTPASS imple-mentation (see below), this approximating function actuallydepends on dual variables Ïk′
t+1, evaluated in iterations k′ =
11 0 0 0 1 k− 1. Thus, more explicitly,
maxyt∈Yt4yt−11 yT 5
vtyt +¯�kt+14yt1 yT 3 Ï
1t+11 0 0 0 1 Ï
k−1t+1 5 (48)
defines the general, primal subproblem for period t and iter-ation k. Slightly different dual subproblems are also solved,as described below.
One might “process” subproblems using a variety ofsequences or methods, and we simply use the FASTPASSmethod identified by Wittrock (1985) as the most effec-tive method among the three tests. (See also Gassman1990.) Specifically, iteration k includes (i) a primal pass,which, for t = 11 0 0 0 1 T , uses yt−1 and the period-t sub-problem to solve for yt , and (ii) a dual pass, which, for
Vossen, Wood, and Newman: Open-Pit Mine Block Sequencing782 Operations Research 64(4), pp. 771–793, © 2016 INFORMS
t = T − 11 0 0 0 12, uses the most recent dual solution Ït+1 tosolve for Ït , which generates a new Benders cut to improvethe approximating function for period t−1. (The dual passinitializes with ÏT taken from the last step in the precedingprimal pass.) Figure 4 illustrates a generic iteration k of thisalgorithm for T = 4; Figure 5(a) gives a condensed illus-tration, which simplifies comparison to other algorithmicvariants. Appendix A provides details on the subproblemssolved in an FBD algorithm, including the elastic formula-tion, an expanded representation of the vector Ït , and therecursive definition of optimality cuts.
3.2. Aggregate Resource Constraints
Computational results in Section 4 show that standard FBDruns slowly. To improve upon these results, we exploitaggregate resource constraints. Note that, in effect, we havealready exploited aggregations of temporal-precedence con-straints (27) to define the bounds y t ¶ yt′ ¶ yt used inYt′4y t1 yt5; see Equation (37).
The idea is simple: total resource consumption from theend of time period t1 to the end of time period t2, t2 > t1,must satisfy both lower and upper bounds on resource con-sumption accumulated from periods t1 +1 through t2. Morespecifically, by summing constraints (26) appropriately andnoting the cancellations that occur because of the station-ary matrix A, we see that Yt′4y t1 yt5 as used in (39) can be
Figure 4. A generic FASTPASS iteration of a standard NBD algorithm that solves the forward recursion of a staircaseLP with T = 4 periods; see constraints (42)–(45).
Start iteration k End iteration k
4.
1.
2.
3.
Notes. The number on the left indicates time period t, the left box represents the primal subproblem for t, and the right box represents the dual subproblemfor t. Downward-pointing arrows correspond to the primal outputs of the subproblems, and upward pointing arrows to dual outputs. The dashed arrow hereand in other figures indicates that no other subproblem uses the indicated output. (That output is saved, however, in case it constitutes part of an optimalprimal solution.) The dual vector Ïk
t represents 4Ákt 1 Â
kt 1 Ä
kt 1 Ã
kt 5 as described in Appendix A.
replaced by the following set of constraints, again for anyt1 t′1 t ∈T0+ with t < t′ < t:
Y +
t′ 4y t1yt5
=
y t¶yt′ ¶ yt
∣
∣
∣
∣
∣
∣
∣
∣
∣
∣
∣
∣
∣
Ht′yt′ ¶01
and A4yt′ − y t5¶t′∑
�=t+1
q� if t′¾ t+1
and A4yt−yt′5¶t∑
�=t′+1
q� if t′¶ t−1
0
(49)
In general, Y +
t′ 4y t1 yt5 ⊆ Yt′4y t1 yt5, with strict inclusionpossible, except that the two constraint sets become identi-cal when t′ = t + 1 = t − 1.
The aggregate constraints in (49) (i.e., the constraintsinvolving the matrix A) link period t′ with periods t and t,thereby enabling the use of (39) to derive more generalproblem recursions. To illustrate, the following subsectionsdescribe three different HBD recursions. Variations on theserecursions could be made theoretically valid for any staircasemodel with cumulative variables, but they seem unlikelyto be computationally attractive without the use of aggre-gate constraints. We also note that the ideas describedbelow have led us to experiment with more direct aggre-gation/disaggregation heuristics for solving MIP and IP
Vossen, Wood, and Newman: Open-Pit Mine Block SequencingOperations Research 64(4), pp. 771–793, © 2016 INFORMS 783
Figure 5. Simplified diagrams for a FASTPASS iteration of NBD for a staircase LP with T = 4.
2.
3.
4.
4.
2.
Start iteration k Start iteration kEnd iteration k End iteration k
(a) (b)
1.
3.
1.
Notes. (a) Represents the standard forward recursion (FBD), which Figure 4 depicts in more detail. (b) Represents an iteration of FBD enhanced withaggregate constraints (FBD-A); see Equations (50)–(52).
approximately; for example, see Van Den Heever and Gross-mann (2000). Thus far we have been unsuccessful, however.
3.3. Forward Nested Decomposition withAggregate Resource Constraints (FBD-A)
The use of aggregate resource constraints, together witha slight reordering of the recursion, can improve FBDsubstantially. Intuitively, by initiating the recursion with amodel that aggregates constraints over the complete timehorizon, the solution process obtains some initial guid-ance from a “weakly constrained UPL solution,” which thestandard forward recursion cannot supply. Specifically, thisrecursion, denoted “FBD-A,” can begin based on period Tto take advantage of the aggregate constraints definedthrough Y +
T 4y01 yT 5:
�∗4y01 yT 5
= maxyT ∈Y+
T 4y01 yT 5vT yT + �14y01yT 5 (50)
= maxyT ∈Y+
T 4y01 yT 5vT yT +
{
maxy1∈Y14y01yT 5
v1y1 + �24y11yT 5}
(51)
000
= maxyT ∈Y+
T 4y01 yT 5vT yT +
{
maxy1∈Y14y01yT 5
v1y1
+
{
· · ·
{
maxyT−1∈YT−14yT−21yT 5
vT−1yT−1
}
· · ·
}}
0 (52)
Figure 5(b) depicts a single FASTPASS iteration of thisrecursion for T = 4. Note that, after the first step to identify
a value for yT , the recursion simply follows the pattern setout in Equations (46) and (47), but with T replacing T .
To gain some insight into the value of the aggregate sub-problem, consider the first subproblem of the first primalpass of FBD-A. (See Equation (50) and Figure 5(b).) Thatsubproblem is
maxyT ∈Y+
T 4y01 yT 5vT yT +
¯�14y01yT 5= maxyT ∈Y+
T 4y01 yT 5vT yT (53)
because ¯�14y01yT 5, the approximation to �14y01yT 5, is null.Given the aggregate constraints, this recursion begins byidentifying a weakly constrained UPL solution and thenidentifies a period-by-period extraction schedule for this“estimated ultimate pit.” (Each iteration of FBD-A thenupdates its estimates of both the aggregate and disaggregatequantities.) Compare that to FBD, which begins by solvingan LP that corresponds to greedily excavating the pit fromperiod 1 to period T : profitable but poor, initial, globalmyopic decisions in early time periods may lead to a poorinitial global solution, from which the algorithm recoversonly at the expense of extra iterations.
3.4. Reverse Nested Decomposition withAggregate Resource Constraints (RBD-A)
For a standard staircase model without cumulative vari-ables, say, a production-inventory problem (e.g., Gabbay1979), a reverse nested decomposition does not seem sensi-ble, especially in early iterations. In particular, in the ordert = T 1T − 11 0 0 0 12, a reverse decomposition would try to
Vossen, Wood, and Newman: Open-Pit Mine Block Sequencing784 Operations Research 64(4), pp. 771–793, © 2016 INFORMS
estimate optimal production quantities in time period t, butwithout having a reasonable approximation of the optimal,total production up to time period t − 1 (i.e., without hav-ing a reasonable approximation of the optimal inventory atthe beginning of period t). Cumulative variables yt in RMIPdo represent total production (extraction) for each block upthrough time period t, but a mirror image in time of thesimple forward recursion (see Equations (42)–(45)) wouldprovide little guidance as to resource consumption in earlyiterations.
By applying aggregate resource constraints, however,we can also obtain global guidance in a reverse recur-sion. Specifically, using (49), the following recursion de-scribes a “reverse decomposition with aggregate constraints”(RBD-A):
�∗4y01 yT 5
= maxyT ∈Y+
T 4y01 yT 5vT yT + �T−14y01yT 5 (54)
= maxyT ∈Y+
T 4y01 yT 5vT yT +
{
maxyT−1∈Y
+T−14y01yT 5
{
vT−1yT−1
+ �T−24y01yT−15}}
(55)
000
= maxyT ∈Y+
T 4y01 yT 5vT yT +
{
maxyT−1∈Y
+T−14y01yT 5
vT−1yT−1
+
{
· · ·
{
maxy1∈Y
+1 4y01y25
v1y1 + �0���>
0
4y01y15}
· · ·
}}
(56)
Figure 6. A single iteration for two versions of nested Benders decomposition with cumulative variables and aggregateconstraints: (a) a reverse decomposition (RBD-A) and (b) a bisection decomposition (BBD-A).
(a) (b)
Start iteration k End iteration k Start iteration k End iteration k
4.
2.
2.
1. 3.
4.
1.
3.
Notes. A FASTPASS processing method, or a generalization thereof, applies to both recursions. Note that Y +
1 4y01 yk
25= Y14y01 yk
25 in both (a) and (b) andthat Y +
3 4yk21 yk
45= Y34yk
21 yk
45 in (b).
More simply, the RBD-A recursion may be summarizedthrough the repeated application of the following relation-ships, starting with t = T and with fixed values t = 0,y0 = y0, and yT = yT :
�∗4yt1yt5 = �t−14y01yt5 (57)
= maxyt−1∈Y
+
t−14y01yt5vt−1yt−1 + �∗4y01yt−150 (58)
Figure 6(a) illustrates a FASTPASS iteration of RBD-Afor T = 4.
3.5. A Tree Decomposition Using AggregateResource Constraints (BBD-A)
We would not use a reverse decomposition without aggre-gate resource constraints, and we would not create a gen-eralizing tree decomposition without them, either. For aparticular application, a user might tailor a recursion tocharacteristics of the modeled system but, for simplicity,we describe “BBD-A,” which defines a “bisection decom-position with aggregate resource constraints.” Also for sim-plicity, we specialize to T = 2n for some integer n> 1:
�∗4y01yT 5= maxyT ∈Y+
T 4y01 yT 5vT yT +�T /24y01yT 5 (59)
= maxyT ∈Y+
T 4y01 yT 5vT yT +
{
maxyT /2∈Y+
T /24y01yT 5vT /2yT /2 +�T /44y01yT /25
+�3T /44yT /21yT 5}
1 (60)
where �T /44y01yT /25= maxyT /4∈Y
+T /44y01yT /25
vT /4yT /4 +�T /84y01yT /45
+�3T /84yT /41yT /251 (61)
Vossen, Wood, and Newman: Open-Pit Mine Block SequencingOperations Research 64(4), pp. 771–793, © 2016 INFORMS 785
Figure 7. Tree representation of nested and hierarchicalBenders decompositions: (a) FBD (b) FBD-A(c) RBD-A (d) BBD-A.
0, 1, 5
1, 2, 5
2, 3, 5
3, 4, 5
(a) (b)
0, 4, 5
0, 1, 4
1, 2, 4
2, 3, 4
0, 4, 5
0, 3, 4
0, 2, 3
0, 1, 2
0, 4, 5
0, 2, 4
0, 1, 2 2, 3, 4
(c) (d)
Note. The numbers in each vertex are 6 t1 t1 t7, with t in a bold font toemphasize that the value for yt is being determined at that vertex.
and �3T /44yT /21yT 5
= maxy3T /4∈Y
+3T /44yT /21yT 5
v3T /4y3T /4 +�5T /84yT /21y3T /45
+�7T /84y3T /41yT 51 (62)
etc.After the first step to identify a value for yT , we may
summarize the recursion for BBD-A through the repeatedapplication of the following relationships, starting with t =
0 and t = T , and with fixed values y0 = y0 and yt = yT :
�∗4yt1yt5 = �t4yt1yt51 where t = 4t + t5/2 (63)
= maxyt∈Y
+t 4yt 1yt5
vtyt + �∗4yt1yt5+ �∗4yt1yt50 (64)
If T is not a multiple of 2, t may be defined as the integerfloor or ceiling of 4t+ t5/2; in fact, computational tests inSections 4 and 6 apply the integer-floor operator.
Figure 6(b) illustrates an iteration of BBD-A for T = 4.We generalize the FASTPASS subproblem processingmethod by (i) viewing the decomposition diagram as a di-rected tree with all arcs pointing downward, away from theroot vertex; (ii) processing primal subproblems in any topo-logical (acyclic) ordering of that tree; and (iii) processingdual problems in a reverse topological ordering of the tree,but skipping leaf vertices. Figure 7(d) shows the diagram ofFigure 6(b) viewed as a tree, while Figures 7(a)–(c) showhow other, simpler decomposition schemes may be viewedas trees, also.
3.6. Hierarchical Benders Decomposition (HBD)
For simplicity, we refer to a tree decomposition that usescumulative variables and aggregate resource constraints as ahierarchical Benders decomposition” (HBD). Both RBD-Aand BBD-A are examples, even though the tree associatedwith RBD-A has an especially restricted structure. FBD-Auses aggregate constraints only in its first stage, but we alsoinclude that as a special case of HBD. On the other hand, it
should be clear that HBD allows even more general recur-sions than those described. For instance, seasonal effectsmight make this decomposition scheme possible over twoyears at a monthly level of detail: two years, one year, onequarter, one month.
One difficulty with standard Benders decomposition isthat early iterations make only slow progress to an opti-mal solution because the few cuts available provide apoor approximation of the subproblem’s or subproblems’contribution to the overall objective function (Geoffrionand Graves 1974). When NBD is used to solve multi-stage stochastic programming problems, several authorshave shown how the application of special “preliminarycuts” can help address this issue (Infanger 1994, pp. 96–99;Morton 1996). In particular, Infanger generates preliminarycuts based on an aggregate, “expected-value model.” Bycontrast, we exploit an aggregate model directly in the Ben-ders recursion rather than through specialized cuts. To thebest of our knowledge, complete aggregate models have notbeen exploited in multistage stochastic programming.
As a final point in this discussion, we note that, in the-ory, HBD could be applied to certain staircase LPs thatlack the stationary matrix A that appears in resource con-straints (26). For example, suppose that RMIP represents aproduction-inventory-distribution model with time-of-yeareffects in production-line yields (e.g., Brown et al. 2001)and that these are represented by replacing constraints (26)with the following:
At4yt − yt−15¶ qt ∀ t ∈T0 (65)
The cancellations that enable creation of aggregate con-straints in (49) no longer apply. But if we define A =
mint∈8t10001t9At , where the minimization is taken element-wise across the matrices At , then Y +
t′ 4y t1 yt5 defined usingthis matrix A is valid, although it may not be as tight aswhen At = A for all t. Intuitively, the weaker constraintswould still give useful guidance to the decomposition pro-vided that the matrices At vary only modestly with t, i.e.,are “nearly stationary.”
4. Computational Tests: Solving RMIPThis section describes computational tests of HBD for solv-ing instances of RMIP. We use all six data sets describedby Cullenbine et al. (2011), which create problem instancesthat cover 10,819, 18,300, and 25,620 blocks, for 1 to 20time periods; the problem’s name indicates the number ofblocks modeled in the data. (Five of these data sets wereused for testing in Section 2.3.2.) Potential solution meth-ods include a simplex algorithm applied to the “monolithicLP” and each of the four variants of nested Benders decom-position: FBD, FBD-A, RBD-A and BBD-A.
A 64-bit workstation with 16 GB RAM and a 3 GHzquad-core Intel processor performs all computations, run-ning under a Windows operating system. A C++ pro-gram generates all LPs, and CPLEX 12.5 (IBM Corp.
Vossen, Wood, and Newman: Open-Pit Mine Block Sequencing786 Operations Research 64(4), pp. 771–793, © 2016 INFORMS
2013) solves those LPs. Default parameters apply except asfollows:
(i) the solver may use at most four threads (Threads = 4);(ii) CPLEX’s barrier algorithm solves the monolithic
LPs (LPMethod = Barrier);(iii) monolithic LPs are limited to 7,200 seconds of
elapsed computation time (TiLim = 7,200);(iv) the dual simplex algorithm solves the
decomposition’s LPs (LPMethod = Dual); and(v) that algorithm emphasizes numerical stability
(NumericalEmphasis = true).Note that the barrier algorithm typically solves an LP
“from scratch” more quickly than does the dual simplexalgorithm, hence (ii). But within a decomposition algorithmwith cuts being added to an LP from one iteration to thenext, the dual simplex algorithm becomes quicker becauseit can exploit standard “warm starts,” hence (iv).
A penalty of 100 dollars/ton, discounted at the model’sstandard rate, applies to the violation of each resource con-straint in period t; this penalty corresponds to p−
rt and p+rt
defined in Appendix A. Also, all decomposition algorithmsenforce a relative optimality tolerance of �RMIP = 10−4.
Problem-specific preprocessing, adjusted for poten-tially fractional blocks, helps reduce model sizes. This
Table 2. BD25620A: Model sizes for instances of RMIP generated as monolithic LPs and sizes for nested subproblemsencountered during solution by decomposition.
Initial Reduced
LP or nested subproblem Variables (num.) Constraints (num.) Variables (num.) Constraints (num.)
RMIP, T = 20 6431424 310081805 4861008 219821412RMIP, T = 5 1281101 6441240 1021072 5911389Avg. FBD subproblem w/o cuts 321040 1291862 271904 1251726Avg. FBD-A subproblem w/o cuts 251623 1091634 71130 21241Avg. RBD-A subproblem w/o cuts 251622 1081353 107 11022Avg. BBD-A subproblem w/o cuts 251623 1091634 11455 71009
Notes. “RMIP, T = 20” corresponds to the largest monolithic LP generated, and “RMIP, T = 5” corresponds to the largest such model that solves in under7,200 seconds. Averages (“Avg.”) are taken across all periods in a 20-period model. “Initial” statistics are generated using problem-specific preprocessing,while “Reduced” statistics are those obtained after applying CPLEX’s “prereduce” methods.
Table 3. BD25620A: Solution statistics for instances of RMIP solved by a barrier algorithm and by decomposition.
Decomposition
Mono-lith FBD FBD-A RBD-A BBD-A
T (sec.) (sub.) (piv.) (sec.) (sub.) (piv.) (sec.) (sub.) (piv.) (sec.) (sub.) (piv.) (sec.)
1 507 — — — — — — — — — — — —2 4806 4 61091 900 4 61773 804 4 61773 804 4 61773 8043 33301 11 121965 1507 7 201302 2602 7 71099 808 7 201302 26034 1124008 16 191854 2208 10 71098 1103 10 71114 1103 9 71050 11025 6132004 29 291830 3305 21 81716 1505 13 81217 1404 19 101736 160810 † 82 761141 8906 46 171510 3408 64 101171 2306 40 111334 220115 † 211 1781743 23503 99 181228 4401 127 171744 4309 103 291631 730820 † 248 1841473 23506 172 161957 5700 324 211825 7601 175 161000 6007
Notes. This table lists solution seconds “(sec.)” for all methods and, for the decomposition algorithms, it lists the number of subproblems solved “(sub.)”and simplex pivots “(piv.).” The symbol “†” indicates that the problem could not be solved within 7,200 seconds. Decomposition algorithms skip thesolution of a subproblem if that subproblem immediately follows the generation of a nonviolated cut.
preprocessing eliminates variables and constraints bybounding the earliest and latest time periods in which ablock can be extracted. (See Lambert et al. 2014 for details,but note that we do not use the “enhanced early starts”described in that paper.) Preprocessing requires only a fewseconds of computation and is performed just once for eachdata set, so we do not report the corresponding computationtimes. Reported times do include CPLEX’s standard “pre-reduce” methods, however.
Initial computation focuses on one of the two largestdata sets, BD25620A. For reference, Table 2 displays somemodel-size statistics for (i) the monolithic LP that is gen-erated for T = 5 and for T = 20 and (ii) the average sub-problem sizes observed while applying the decompositionalgorithms for T = 20. Table 3 shows solution times overa range of values for T .
General trends seen for BD25620A in Table 3 hold forall data sets, so we use these results to cull the clearlyinferior methods, which are solution as a monolithic LPand solution via FBD. In particular, the barrier algorithmapplied to the monolith cannot compete with FBD, but FBDcannot compete with the hierarchical decomposition meth-ods. (When T = 3, BBD-A does perform poorly, and FBDis faster, but this is the only such case for BD25620A.)
Vossen, Wood, and Newman: Open-Pit Mine Block SequencingOperations Research 64(4), pp. 771–793, © 2016 INFORMS 787
The smaller subproblem sizes shown in Table 2 for HBDcompared to FBD result from HBD’s aggregate constraintsenabling the fixing of more variables to zero or one, andthese statistics do help explain the computational superior-ity of HBD methods over FBD. Other problem instances donot show such stark differences, but in no case does FBDoutperform any HBD method.
We have experimented with several methods to improvethe computational performance of FBD, but with no suc-cess. First, neither the primal simplex algorithm nor thebarrier algorithm in CPLEX is faster than the dual sim-plex algorithm for solving the relevant LPs. Second, ourattempts at “stabilizing” primal and/or dual solutions bysolving primal and/or dual subproblems using CPLEX’sbarrier algorithm all fail. In particular, the use of interior-point solutions can improve the empirical convergenceof decomposition algorithms (e.g., Rousseau et al. 2007,Singh et al. 2009), but the extra time required to findsuch solutions produces at best minor reductions in thenumber of major iterations for FBD while greatly increas-ing solution times. With some significant implementationaleffort, other specialized techniques might yield computa-tional benefits (Ruszczynski 1986, Gondzio et al. 1997,Elhedhli and Goffin 2004). However, as seen in Table 3,the use of aggregate constraints to create FBD-A producesdefinite computational benefits; furthermore, the implemen-tational effort for this “specialized technique” is modest.
Table 4 shows the results for all six data sets usingthe solution methods that pass the “culling test” describedabove. Solution times remain reasonable for all hierarchi-cal decompositions, even as the number of time periodsbecomes large. This suggests that the technique commonto all, namely, the use of aggregate constraints, is key tocomputational efficiency. We postpone further discussion ofthese computational results until Section 7.
5. Combining HBD and HeuristicMethods to Solve MIP
Computational tests above show that HBD can solvemedium-sized instances of RMIP that have both upper- andlower-bounding resource constraints, but OPBS requiresMIP-valid solutions for MIP. This section describes aheuristic that combines HBD and branch and bound to pro-duce high-quality solutions for MIP to all of the test prob-lems solved as LPs above.
As in Section 3.5, a graph G = 4V 1E5 describes thehierarchy tree, with vertices i ∈ V labeled in the orderof a primal pass for HBD. We adjust the notation sothat RMIPt4i54y
∗
t4i51 y∗
t4i55, which has solution yt4i5, denotesthe LP subproblem at vertex i in the tree. The notationMIPt4i54y t4i51 yt4i55 indicates the corresponding MIP (with-out explicit binary variables), which has MIP-valid solu-tion y∗
t4i5. The following describes the complete heuristicfor solving MIP.
Table 4. Solution times, in seconds, for 10k-, 18k-, and25k-block instances of RMIP solved usingHBD.
BD10819A BD10819F
T FBD-A RBD-A BBD-A FBD-A RBD-A BBD-A
2 008 008 008 008 008 0083 104 105 104 106 106 1064 203 205 203 203 205 2055 304 308 309 303 307 30510 1607 1706 1608 1704 1901 150915 4504 4504 3702 4409 4600 350720 8802 9705 6600 10409 9403 5500
BD18300A BD18300B
T FBD-A RBD-A BBD-A FBD-A RBD-A BBD-A
2 1901 1901 1901 102 101 1013 205 204 205 108 107 1074 303 302 305 503 503 5025 401 1806 406 1000 803 30110 1103 1600 1600 1005 909 140515 11407 3404 2607 1808 2603 270120 3600 5606 4103 4705 6900 3206
BD25620A BD25620B
T FBD-A RBD-A BBD-A FBD-A RBD-A BBD-A
2 804 804 804 700 700 7003 2602 808 2603 905 909 9054 1103 1103 1102 1303 1302 13025 1505 1404 1608 1600 1600 150610 3408 2306 2201 6605 5903 630015 4401 4309 7308 16004 9908 700420 5700 7601 6007 25102 57407 19803
Procedure HMIPH (“Hierarchical MIP Heuristic”)Input: Full problem data for MIP, relative optimality
tolerance �SUB > 0 for subproblem MIPs.Output: A MIP valid solution to MIP, 4y∗
11 0 0 0 1 y∗
T 5.Notation: yt (y∗
t ) denotes an LP (MIP-valid) solution of amodel for period t.
{Step 0: Using a specific version of HBD, solve RMIP
for 4y1,…,yT 5;For (i = 1 to T ) { /∗ In the order of a primal pass of
the decomposition tree ∗/
Step 1: If (i 6= 1) {Beginning with cuts generated in step 0, andgenerating new cuts as necessary, use the sameversion of HBD to solve RMIPt4i54y
∗
t4i51 y∗
t4i55
for yt4i5;/∗ yt4i5 serves as a starting point to solveMIPt4i54y
∗
t4i51 y∗
t4i55. The MIP-valid inputs y∗
t4i5
and y∗
t4i5 are available because of thevertex-processing order. ∗/
}
Vossen, Wood, and Newman: Open-Pit Mine Block Sequencing788 Operations Research 64(4), pp. 771–793, © 2016 INFORMS
/∗ The next step solves a “MIP subproblem.” ∗/Step 2: With relative optimality tolerance �SUB1 solve
MIPt4i54y∗
t4i51 y∗
t4i55 for y∗
t4i5 using the branch-and-bound procedure outlined in the textbelow;
9Print(“The MIP-valid, approximate solution toMIP is,” 4y∗
11 0 0 0 1 y∗
T 5);}
The branch-and-bound procedure used in step 2 resem-bles the “SOS2 constraint branching” of Beale and Tom-lin (1970). Specifically, if at vertex i the solution yt4i5to RMIPt4i54y
∗
t4i51 y∗
t4i55 does not satisfy integrality require-ments for an implicitly determined xt4i5, blocks b and bmust exist such that ybt4i5 > 0, but ybt4i5 < 1. That is, someblock b is partially extracted in the current solution, yetblock b directly above b is not completely extracted. Whenthis happens, we branch by enforcing “ybt = 0” or “ybt = 1.”(This does not enforce a partition of MIP’s feasible regionin terms of y, but it does correspond to a partition ofthe full feasible region defined in terms of 4x1y5.) Nodesin the branch-and-bound tree are always feasible and thusare fathomed by bound. The following three rules controlbranching:
1. Choose the next node for branching based on a “best-bound criterion” (e.g., Linderoth and Savelsbergh 1999).
2. Given the branching node i, among blocks b suchthat ybt4i5 > 0 and ybt4i5 < 1, branch on the block b
˜
havingthe largest number of fractional, direct predecessors andsuccessors. (Block b′ is a direct successor of b if b is adirect predecessor of b′.)
3. Branch first in a direction analogous to “branch up,”by enforcing this restriction: yb
˜
t = 1. (The complementarybranching direction is thus yb
˜
t = 0.)HMIPH is not an exact algorithm because branch and
bound applies only to individual MIP subproblems in thehierarchy. But the always-feasible, MIP-valid, final solu-tion 4y∗
11 0 0 0 1 y∗
T 5 defines a lower bound on �∗MIP, and the
optimal objective value for MIPt4154y∗
t4151 y∗
t4155 defines anupper bound, a bound that may be better than �∗
RMIP.
6. Computational Results: Solving MIPBased on the instances of RMIP already investigated (seeTable 4), this section evaluates the computational perfor-mance of HMIPH for solving MIP, i.e., the full MIP modelfor OPBS. Settings for LP subproblems in steps 0 and 1of the procedure remain as in Section 4; MIP subprob-lems in step 2 incorporate a relative optimality toleranceof �SUB = 5 × 10−4. (The values for �SUB and �RMIP areselected together based on empirical tests, which show thepair to yield a good trade-off between solution speed and“accuracy,” that is, observed optimality gap.)
Table 5 shows the results for all six instances of MIPusing HMIPH with the three versions of HBD. Solutiontimes remain reasonable. For example, based on BBD-A,
each 20-period instance of MIP solves in less than twicethe time required to solve the corresponding instance ofRMIP. Moreover, the computed relative optimality gapsindicate that the heuristic consistently yields high-qualitysolutions. We also note that any resource-constraint viola-tions are negligible.
The bisection decomposition BBD-A leads to good solu-tions to MIP more quickly than do the other decomposi-tions. The next section indicates why this may be true.
7. Discussion of Computational Resultsfor Both RMIP and MIP
The results presented in Table 3 indicate the superiorityof all HBD variants over a standard implementation ofnested Benders decomposition (FBD) for solving RMIP:the HBD variants are usually faster than FBD, and whenT ¾ 10, they are always two to five times faster. No cleartrend appears in Table 4, however, to indicate a clearcomputational winner among the HBD variants. But itshould be easy to parallelize BBD-A and, with T /2 pro-cessors and sufficient computer memory, total solution timemight reduce by a factor approaching 4logT 5/T . Thus, webelieve that, of the HBD variants, BBD-A holds the great-est promise for solving staircase LPs.
Although no HBD method is clearly faster than anyother for solving RMIP, Table 5 shows that BBD-A doesgain a computational advantage when used to help solvethe largest instances of MIP. Specifically, applications ofHMIPH based on BBD-A produce solution times on the20-period instances that range from 47% to 94% of thenearest rival, with an average of 77%. Average subproblemsizes for RMIP based on BBD-A are not smaller than forthe other HBD variants (see Table 2), so we attribute mostof this advantage to the more “balanced branching” thatmust take place in HMIPH when basing computations onBBD-A.
More specifically, well-balanced branch-and-bound enu-meration avoids branching such that one side of the branch-ing partition is strongly restricted while the other sideis not. Balanced branching explains much of the effi-ciency improvements seen with (i) branching based on spe-cial ordered sets (Beale and Tomlin 1970, Hummeltenberg1984), (ii) the implicit-constraint branching exploited byRyan and Foster (1981) to help solve set-partitioning prob-lems, and (iii) the explicit-constraint branching exploited byAppleget and Wood (2000) to help solve certain binary IPs.
To give a simple, intuitive example, suppose that basedon FBD-A or BBD-A, HMIPH is applied to an all-binary,T -period instance of OPBS with variables xbt . For eithervariant, at vertex i = 1, HMIPH determines xT , whichspecifies for each block b whether or not the block isextracted over the full time horizon. At vertex i = 2,HMIPH based on FBD-A would select a fractional variablexb1 and branch as follows: (i) block b is extracted in periodt = 1, or (ii) if it is extracted, block b is extracted in some
Vossen, Wood, and Newman: Open-Pit Mine Block SequencingOperations Research 64(4), pp. 771–793, © 2016 INFORMS 789
Table 5. Solution statistics for 10k-, 18k- and 25k-block instances of MIP solved using HMIPH.
BD10819A BD10819F
Gap (%) Soln. time (sec.) Gap (%) Soln. time (sec.)
T FA RA BA FA RA BA FA RA BA FA RA BA
2 008 008 008 202 202 203 008 008 008 203 203 2043 007 007 007 301 301 301 008 007 008 302 302 3024 006 006 006 408 406 403 008 007 007 407 408 4085 007 006 006 509 600 601 007 006 006 506 601 60110 005 005 005 2309 2206 2406 005 005 005 2502 2406 220315 006 006 006 6001 5808 5209 005 005 005 5907 5800 480220 005 006 006 12101 12003 9602 005 006 006 13507 11707 8905
BD18300A BD18300B
Gap (%) Soln. time (sec.) Gap (%) Soln. time (sec.)
T FA RA BA FA RA BA FA RA BA FA RA BA
2 002 002 002 2501 2407 2408 000 000 000 102 102 1023 004 004 004 907 1200 906 000 000 000 200 200 2004 004 004 004 900 808 804 000 000 000 507 509 5065 003 003 003 1005 2506 905 000 000 000 1104 904 40410 002 002 002 2101 2903 2904 000 000 000 1408 1406 180515 002 002 002 13800 8202 5804 000 000 000 3803 3903 350320 002 002 002 8300 12806 7802 000 000 000 7600 11009 5408
BD25620A BD25620B
Gap (%) Soln. time (sec.) Gap (%) Soln. time (sec.)
T FA RA BA FA RA BA FA RA BA FA RA BA
2 000 000 000 809 805 805 000 000 000 1008 1006 10063 000 000 000 2607 903 2609 001 001 001 2407 2107 24044 000 000 000 1203 1109 1200 002 002 002 2709 2501 25065 000 000 000 1704 1503 1801 004 004 005 5304 4806 500110 000 000 000 4404 2900 3109 102 102 102 19702 13008 2890615 000 000 000 6102 6803 9000 103 104 102 85304 26604 1580320 000 000 000 9603 15400 8602 102 102 101 85206 1144104 40306
Note. “FA” = FBD-A, “RA” = RBD-A, “BA” = BBD-A, and “Gap” = relative optimality gap.
period t > 1. The first element in this partition is stronglyrestricted while the second is only weakly restricted, sothis branching scheme is unbalanced. By contrast, at vertexi = 2, HMIPH based on BBD-A would identify a fractionalvariable xb1�T /2�, and branching would execute this better-balanced partition: (i) block b is extracted in the first halfof the time horizon, or (ii) if it is extracted, block b isextracted in the second half of the time horizon.
For computational evidence of balanced branching, con-sider the most computationally challenging problem in-stance, BD25620B. BBD-A and FBD-A have about thesame number of violated integrality restrictions whenbranching commences at each decomposition-tree vertex,about 950 when summed over all i. But BBD-A createsa total of only 98 branch-and-bound nodes across all MIPsubproblems while FBD-A creates 369. (RBD-A is worsethan FBD-A on both measures.)
As a final point on computation, note that HMIPH canbe applied to solve IP rather than MIP. Steps 0 and 1 ofHMIPH remain the same because the solution to RMIP is
also the solution to the LP relaxation of IP. But then, step 2uses branch and bound to solve the integer-programminganalog of RMIPt4i54y
∗
t4i51 y∗
t4i55. We have experimented withsuch a heuristic, but with little success. Specifically, whilethe heuristic should yield an elastically feasible solution,the lack of flexibility in the model formulation yields indi-vidual integer-programming subproblems in step 2 that,typically, do not solve in an hour of computation time.
8. Conclusions and Future ResearchThis paper has described a new mixed-integer-programmingmodel for an open-pit block sequencing problem andhas developed a special decomposition procedure for thatmodel’s solution. Unlike most other work on OPBS, our for-mulation MIP incorporates lower bounds on resource con-sumption in each period in addition to the standard upperbounds. Uniquely, this formulation also allows for fractionalblock extraction in a mine while still satisfying pit-wallslope restrictions.
Vossen, Wood, and Newman: Open-Pit Mine Block Sequencing790 Operations Research 64(4), pp. 771–793, © 2016 INFORMS
The staircase constraint structure of MIP enables a new“hierarchical” Benders decomposition for solving MIP’slinear-programming relaxation “RMIP.” HBD generalizesnested Benders decomposition, taking advantage of twospecial techniques: it formulates the model using cumu-lative variables and adds time-aggregated resource con-straints that provide useful guidance to the overall solutionprocedure.
Computational testing in this paper does not exploit par-allel solution of nested subproblems, but we believe thatthis will be a fruitful line of work to follow. A standardimplementation of a forward recursion in NBD leaves lit-tle room for parallel computation because the kth primalpass of approximate subproblems RMIPk
t solves those inthe order t = 11 0 0 0 1 T , and a dual pass solves them in thereverse order. But HBD includes “tree decompositions” inwhich a subproblem RMIPt decomposes into RMIPt1
andRMIPt2
such that approximating subproblems RMIPkt1
andRMIPk
t2could be solved in parallel, both in primal and dual
passes. In fact, given �T /2� processors, it may be possibleto reduce solution time for RMIP by a factor approaching4logT 5/T .
We also note that NBD has been used extensively forsolving multistage stochastic programs, so the usefulnessof HBD needs to be explored for such applications.
Acknowledgments
The second author thanks the Office of Naval Research and theAir Force Office of Scientific Research for support of his research.
Appendix A. Solving RMIP with a ForwardNested Benders Decomposition, FBD
This appendix outlines how standard nested Benders decomposi-tion solves a variant of RMIP in Section 2.2 using the forwardrecursion (FBD), as described by Equations (42)–(45) in Sec-tion 3.1.
The variant of RMIP allows penalized violation of resourceconstraints (26) using elastic variables
st = 44s−
1t1 0 0 0 1 s−
�R�t51 4s+
1t1 0 0 0 1 s+
�R�t55>1 (A1)
where s−rt and s+
rt correspond to violations of the lower-boundingand upper-bounding constraints (5) and (6), respectively. The cor-responding vector of constraint-violation penalties, in units of dol-lars/ton, is
pt = 44p−
1t1 0 0 0 1 p−
�R�t51 4p+
1t1 0 0 0 1 p+
�R�t550 (A2)
Extending the model (46)–(47) with elastic resource constraints,the following recursively defines the full model and the cost-to-gofunction �t4yt−11 yT 5. Given y0 and yT , for t = 11 0 0 0 1 T ,
RMIPt4yt−11 yT 52 �t4yt−11 yT 5
= max0¶yt¶yT 1 st¾0
vtyt −ptst + �t+14yt1 yT 5 (A3)
s0t0 Ayt − Ist ¶ qt +Ayt−1 (A4)
−Iyt ¶−I yt−1 (A5)
Htyt ¶ 01 (A6)
where �T 4yT−11yT 5 ≡ 0, and the constraints −AyT ¶ qT − AyTimplied by (37) are omitted because they are constructed so as tobe vacuous. (See the definition of qt under “Additional Notation”in Section 3. Note that constraints analogous to these do appearin BBD-A, namely, constraints (B2), provided that t 6= T .)
Theorem 2. FBD, as described in Section 3.1 for solving RMIP,converges when applied to the version of RMIP (A3)–(A6) thatreplaces standard resource constraints with elastic ones.
Proof. (For reference, (25)–(31) directly define RMIP in thetext, and (42)–(47) define that model recursively.) Viewing yt−1 =
yt−1 as a parameter vector, RMIPt4yt−11 yT 5 above may be rewrit-ten as
RMIPt4yt−11 yT 52
�t4yt−11 yT 5= maxyt∈Y 4yt−11 yT 5
vtyt −pt4A4yt − yt−15−qt5+
+ �t+14yt1 yT 51 (A7)
where Yt4yt−11 yT 5 ≡ 8yt � yt−1 ¶ yt ¶ yT 1Htyt ¶ 09. Letting t =
t − 1 and t = T , we see that RMIPt4yt−11 yT 5 simply relaxesYt+14yt1 yT 5 in the formulation (46)–(47) and also replaces thepiecewise-linear concave function of yt , �t+14yt1 yT 5, with a dif-ferent piecewise-linear concave function of yt . Thus, standardconvergence theory holds. �
The variables st are auxiliaries that help us compute approxi-mations to a revised piecewise-linear function, but the model canbe stated without them. The revised formulation ensures feasibilityof the period-t subproblem provided that that y0 ¶ yt ¶ yT , whichis guaranteed, so only Benders optimality cuts need be generated.
FBD replaces �t+14yt1 yT 5 in RMIPt4yt−11 yT 5 for each itera-tion k with a piecewise-linear, concave, upper approximation:
¯�kt+14yt1 yT 5≡ mink′=110001k−1
hk′
t + gk′
t yt1 (A8)
with the cut coefficients hk′
t and gk′
t defined in Equations (A14)and (A15) below. Making the replacement yields the following“approximate (nested) subproblem,” with dual variables for therelevant constraints shown in brackets:
RMIPkt 4yt−11yT 52
¯�kt 4yt−11yT 5
≡ maxyt¾01st¾01 ¯�t+1
vtyt −ptst +¯�t+1 (A9)
s.t. Ayt −Ist¶qt +Ayt−1 6Ákt 7 (A10)
−Iyt¶−I yt−1 6Âkt 7 (A11)
Htyt¶0 6Äkt 7 (A12)
−gk′
t yt +¯�t+1¶hk′
t
for k′=110001k−11 6�kk′
t 7 (A13)
where ¯�T ≡ 0 and cuts (A13) are omitted for t = T . Note that(i) Ák
t , Âkt and Äk
t are vectors while �kk′
t is not; (ii) the genericdual variables Ïk
t in the body of the paper now correspond to4Ák
t 1Âkt 1Ä
kt 1Ã
kt 5, where we do now define the vector Ãk
t = 4�k1t 1
0 0 0 1 �kkt 5; and (iii) the vector Äk
t is not actually used in the cut-creation process because the right-hand side of the relevant con-straints is 0.
Vossen, Wood, and Newman: Open-Pit Mine Block SequencingOperations Research 64(4), pp. 771–793, © 2016 INFORMS 791
Finally, we recursively define cut coefficients as follows:
gkt−1 = ÁktA−Âk
t I for t=T 1 T −11000121 (A14)
hkt−1 = Ák
t qt +
k−1∑
k′=1
Ãkk′
t hk′
t for t=T 1 T −11000120 (A15)
FBD adds cuts (A13) dynamically, alternating between a primalpass and a dual pass. In this forward decomposition, the kth pri-mal pass solves each approximate subproblem RMIPk
t 4yt−11 yT 5for ykt in the order t = 11 0 0 0 1 T , given y0 ≡ 0. Next, in the ordert = T 1 0 0 0 12 and using values yt obtained in the most recentprimal pass, the kth dual pass solves RMIPk
t 4yt−11 yT 5 for dualvariables and adds cuts to update the approximating functions¯�kt 4yt−11 yT 5. (The last step of the primal pass, i.e., when t = T ,actually implements the first step of the dual pass.)
Given elastic resource constraints, the kth primal pass yieldsa primal feasible solution to RMIP, namely, 4yk11 0 0 0 1 y
kT 5. Thus,
�k ≡∑T
t=1 vt ykt ¶ �∗. Because ¯�k14y01 yT 5¾ �∗, testing ¯�k14y01 yT 5
− maxk �k ¶ �RMIP, for some �RMIP > 0, yields an appropriate
stopping criterion for the decomposition algorithm.
Appendix B. Solving RMIP with HierarchicalBenders Decomposition, BBD-A
This appendix describes the implementation of BBD-A, a bisec-tion version of HBD that solves an elastic version of RMIP.
Assume that time periods t < t < t are given, with
t = f 4t1 t5=
{
T 1 if t = 0 and t = T ;
�4t + t5/2�1 otherwise.
Also, assume the existence of solution estimates y t and yt , withy0 ¶ y t ¶ yt ¶ yT . As in Appendix A, we solve a revised modelwith elastic resource constraints. In general, two groups of con-straints must be elasticized, however, so we (i) define st and stanalogous to st (see (A1)) and (ii) define pt and pt analogous topt (see (A2)).
Expanding on Equations (63) and (64), the following recursiondefines the LP that is solved through BBD-A.
RMIPt4y t1yt52 �t4y t1yt5
= maxyt¾01 st 1 st¾0
vtyt −ptst −pt st +�t′ 4y t1yt5+�t′′ 4yt1yt5 (B1)
s0t0 Ayt −I st¶t∑
�=t
q� +Ay t (B2)
−Ayt −I st¶t∑
�=t+1
q� −Ayt (B3)
−Iyt¶−I y t (B4)
Iyt¶ I yt (B5)
Htyt¶01 (B6)
where t′ = f 4t1 t5 and t′′ = f 4t1 t5. Of course, the recursion beginswith t = t0 and t = T = T + 1 and is not carried further whent = t + 1.
In the most general case, an iteration k of HBD identi-fies cut coefficients hk′
tt , gk′
tt, gk
′
tt , hk′
tt , gk′
tt , and gk′
tt for k′ =
11 0 0 0 1 k − 1 (see Equations (B17)–(B19), below) and replaces
�t′ 4y t1yt5 and �t′′ 4yt1 yt5. with these piecewise-linear, concave,upper approximations
¯�kt′ 4y t1yt5 = mink′=110001k−1
8hk′
tt + gk′
tt y t + gk′
ttyt9 and (B7)
¯�kt′ 4yt1 yt5 = mink′=110001k−1
8hk′
tt + gk′
tt yt + gk′
tt yt91
respectively. (B8)
Note that because t′ is determined by 6 t1 t7 and t′′ is determinedby 6t1 t7, t′ and t′′ are omitted in defining (B7) and (B8). Alsonote that ¯�kt′ 4y t1yt5≡ 0 if t = t+1 and ¯�kt′′ 4yt1 yt5≡ 0 if t = t−1.In the decomposition’s kth iteration, the approximating model forRMIPt4y t1 yt5 is thus
RMIPkt 4y t1 yt52 �kt 4y t1 yt5
= maxyt 1 st 1 st¾0
vtyt − ptst − pt st +¯�tt +
¯�tt (B9)
s0t0 Ayt − I st ¶t∑
�=t
q� +Ay t 6Áktt7 (B10)
−Ayt − I st ¶t∑
�=t+1
q� −Ayt 6Áktt7 (B11)
− Iyt ¶−I y t 6Âktt7 (B12)
Iyt ¶ I yt 6Âktt7 (B13)
Htyt ¶ 0 6Äkt 7 (B14)
− gk′
ttyt +¯�tt ¶ hk′
tt + gk′
tt y t
for k′= 11 0 0 0 1 k− 1 6�kk′
tt 7 (B15)
− gk′
tt yt +¯�tt ¶ hk′
tt + gk′
tt yt
for k′= 11 0 0 0 1 k− 1 6�kk′
tt 7 (B16)
Note that (i) Ákt t , Á
ktt , Â
kt t , Â
ktt and Äk
t are vectors, while �kk′
tt
and �kk′
tt are not; (ii) the generic dual vectors Ïkt in the text (see
Figure 6(b)) correspond here to 4Ákt t1Á
ktt1Â
kt t1Â
ktt1Ä
kt 1Ã
ktt1Ã
ktt5,
where we do now define vectors Ãktt = 4�k1
tt 1 0 0 0 1 �kk′
tt 5 and Ãktt =
4�k1tt 1 0 0 0 1 �
kk′
tt 5; and (iii) the vectors Äkt remain unused as in FBD
and any other decomposition.As before, we recursively define cut coefficients for (B15) and
(B16):
hktt = Ák
tt
t∑
�=t
q� +Áktt
t∑
�=t+1
q� +
k−1∑
k′=1
�kk′
tt hk′
tt +
k−1∑
k′=1
�kk′
tt hk′
tt 1 (B17)
gktt = ÁkttA−Âk
ttI+
k−1∑
k′=1
�kk′
tt gk′
tt 1 (B18)
gktt = −ÁkttA+Âk
ttI+
k−1∑
k′=1
�kk′
tt gk′
tt 1 (B19)
where these formulas must be applied in a dual-pass order in thedecomposition tree, with leaf vertices omitted.
We do not provide a formal proof that BBD-A convergesbecause it may be viewed as a variant on FBD-A, which doesconverge. To see this equivalence, suppose that BBD-A performsprimal passes in level order and dual passes in the reverse order.This means that all subproblems at the same depth in the treeare solved sequentially. But because the individual subproblems
Vossen, Wood, and Newman: Open-Pit Mine Block Sequencing792 Operations Research 64(4), pp. 771–793, © 2016 INFORMS
within a level are effectively independent, we could solve themsimultaneously, as part of a single “level subproblem.” Supposewe take this view and modify the decomposition algorithm toapply a single cut for each level subproblem. The “new algorithm”is really just FBD-A applied to a model that has been rearrangedinto O4logT 5 stages. Because BBD-A derives from this new algo-rithm by using a cut for each independent subproblem, we seethat BBD-A is just a multicut version of FBD-A.
References
Ahuja R, Magnanti T, Orlin J (1993) Network Flows (Prentice Hall, Engle-wood Cliffs, NJ).
Ahuja RK, Hochbaum DS, Orlin JB (2003) Solving the convex cost inte-ger dual network flow problem. Management Sci. 49(7):950–964.
Akaike A, Dagdelen K (1999) A strategic production scheduling methodfor an open pit mine. Dardano C, Francisco M, Proud J, eds. Proc.28th Internat. Sympos. Comput. Appl. Minerals Industries, APCOM’99 (Colorado School of Mines, Golden, CO), 729–738.
Appleget JA, Wood RK (2000) Explicit-constraint branching for solv-ing mixed-integer programs. Laguna M, González-Velarde JL, eds.Computing Tools for Modeling, Optimization and Simulation (KluwerAcademic Publishers, Boston), 245–261.
Barnhart C, Johnson E, Nemhauser G, Savelsbergh M, Vance P (1998)Branch-and-price: Column generation for solving huge integer pro-grams. Oper. Res. 46(3):316–329.
Beale EML, Tomlin JA (1970) Special facilities in a general mathematicalprogramming system for non-convex problems using ordered sets ofvariables. Lawrence J, ed. Proc. 5th IFORS Conf. (Tavistock, Lon-don), 447–454.
Benders JF (1962) Partitioning procedures for solving mixed-variablesprogramming problems. Numerische Mathematik 4(3):238–252.
Bienstock D, Zuckerberg D (2010) Solving LP relaxations of large-scale precedence constrained problems. Eisenbrand F, Shepherd FB,eds. Proc. 14th Internat. Conf. Integer Programming Combinato-rial Optimization, IPCO ’10, Lecture Notes Comput. Sci., Vol. 6080(Springer, Berlin), 1–14.
Birge JR (1997) State-of-the-art survey—stochastic programming: Com-putation and applications. INFORMS J. Computing 9(2):111–133.
Birge JR, Louveaux FV (1988) A multicut algorithm for two-stagestochastic linear programs. Eur. J. Oper. Res. 34(3):384–392.
Boland N, Dumitrescu I, Froyland G (2008) A multistage stochastic pro-gramming approach to open pit mine production scheduling withuncertain geology. http://www.optimization-online.org/DB_HTML/2008/10/2123.html.
Boland N, Dumitrescu I, Froyland G, Gleixner A (2009) LP-based dis-aggregation approaches to solving the open pit mining productionscheduling problem with block processing selectivity. Comput. Oper.Res. 36(4):1064–1089.
Brown G, Keegan J, Vigus B, Wood K (2001) The Kellogg Companyoptimizes production, inventory and distribution. Interfaces 31(6):1–15.
Caccetta L, Hill S (2003) An application of branch and cut to open pitmine scheduling. J. Global Optim. 27(2–3):349–365.
Cai W (2001) Design of open-pit phases with consideration of scheduleconstraints. Xie H, Wang Y, Jiang Y, eds. Proc. 29th Internat. Sym-pos. Comput. Appl. Minerals Industries, APCOM ’01 (A.A. Balkema,Lisse, Netherlands), 217–221.
Chicoisne R, Espinoza D, Goycoolea M, Moreno E, Rubio E (2012) Anew algorithm for the open-pit mine scheduling problem. Oper. Res.60(3):517–528.
Cullenbine C, Wood K, Newman A (2011) A sliding time window heuris-tic for open pit mine block sequencing. Optim. Lett. 88(3):365–377.
Dagdelen K, Johnson T (1986) Optimum open pit mine productionscheduling by Lagrangian parameterization. Proc. 19th Internat. Sym-pos. Comput. Appl. Minerals Industries APCOM ’86 (SME, Littleton,CO), 127–141.
Dantzig GB, Wolfe P (1960) Decomposition principle for linear programs.Oper. Res. 8(1):101–111.
Denby B, Schofield D (1994) Open-pit design and scheduling by use ofgenetic algorithms. Trans. Institution Mining Metallurgy Section A.Mining Indust. 103:A21–A26.
Eivazy H, Askari-Nasab H (2012) A mixed integer linear programmingmodel for short-term open pit mine production scheduling. MiningTech. 121(2):97–108.
Elhedhli S, Goffin J-L (2004) The integration of an interior-point cuttingplane method within a branch-and-price algorithm. Math. Program-ming 100(2):267–294.
Entriken R (1989) The parallel decomposition of linear programming.Doctoral dissertation, Stanford University, Stanford, CA.
Entriken R (1996) Parallel decomposition: Results for staircase linear pro-grams. SIAM J. Optim. 6(4):961–977.
Epstein R, Goic M, Weintraub A, Catalán J, Santibáñez P, Urrutia R,Cancino R, Gaete S, Aguayo A, Caro F (2012) Optimizing long-termproduction plans in underground and open-pit copper mines. Oper.Res. 60(1):4–17.
Espinoza D, Goycoolea M, Moreno E, Newman A (2012) MineLib: Alibrary of open pit mining problems. Ann. Oper. Res. 206(1):1–22.
Gabbay H (1979) Multi-stage production planning. Management Sci.25(11):1138–1148.
Gassman H (1990) MSLiP: A computer code for the multistage stochasticlinear programming problem. Math. Programming 47(1–3):407–423.
Geoffrion AM, Graves GW (1974) Multicommodity distribution systemdesign by Benders decomposition. Management Sci. 20(5):822–844.
Gershon M (1983) Optimal mine production scheduling: Evaluation oflarge scale mathematical programming approaches. Internat. J. Min-ing Engrg. 1(4):315–329.
Gershon M (1987) Heuristic approaches for mine planning and productionscheduling. Internat. J. Mining Geological Engrg. 5(1):1–13.
Gholamnejad J, Moosavi E (2012) A new mathematical programmingmodel for long-term production scheduling considering geologicaluncertainty. J. Southern African Inst. Mining Metallurgy 112(2):77–81.
Glassey CR (1971) Dynamic linear programs for production scheduling.Oper. Res. 19(1):45–56.
Glassey CR (1973) Nested decomposition and multi-stage linear pro-grams. Management Sci. 20(3):282–292.
Gleixner A (2008) Solving large-scale open pit mining productionscheduling problems by integer programming. Master’s thesis, Tech-nische Universität Berlin, Berlin, Germany.
Gondzio J, Sarkissian R, Vial J-P (1997) Using an interior point methodfor the master problem in a decomposition approach. Eur. J. Oper.Res. 101(3):577–587.
Ho J, Manne A (1974) Nested decomposition for dynamic models. Math.Programming 6(1):121–140.
Hochbaum D, Chen A (2000) Performance analysis and best implemen-tations of old and new algorithms for the open-pit mining problem.Oper. Res. 48(6):894–914.
Hummeltenberg W (1984) Implementations of special ordered sets in MPsoftware. Eur. J. Oper. Res. 17(1):1–15.
IBM Corp. (2013) IBM ILOG CPLEX Optimization Studio CPLEXUser’s Manual, Version 12 Release 5. http://pic.dhe.ibm.com/infocenter/cosinfoc/v12r5/topic/ilog.odms.studio.help/pdf/usrcplex.pdf.
IBM Corp. (2014) IBM ILOG CPLEX Optimization Studio CPLEXUser’s Manual, Version 12 Release 6. http://pic.dhe.ibm.com/infocenter/cosinfoc/v12r6/topic/ilog.odms.studio.help/pdf/usrcplex.pdf.
Infanger G (1994) Planning Under Uncertainty: Solving Large-ScaleStochastic Linear Programs (Boyd & Fraser, Danvers, MA).
Johnson T (1968) Optimum open pit mine production scheduling. Doctoraldissertation, University of California, Berkeley, Berkeley.
Kallio MJ, Porteus EL (1977) Decomposition of arborescent linear pro-grams. Math. Programming 33(1):348–356.
Karlsson J, Rönnqvist M, Bergström J (2003) Short-term harvest planningincluding scheduling of harvest crews. Internat. Trans. Oper. Res.10(5):413–431.
Kawahata K (2007) A new algorithm to solve large scale mine produc-tion scheduling problems by using the Lagrangian relaxation method.Doctoral dissertation, Colorado School of Mines, Golden, CO.
Vossen, Wood, and Newman: Open-Pit Mine Block SequencingOperations Research 64(4), pp. 771–793, © 2016 INFORMS 793
Krige D (1951) A statistical approach to some basic mine valuation prob-lems on the Witwatersrand. J. Chemical Metallurgical Mining Soc.South Africa 52(6):119–139.
Lambert WB, Newman AM (2014) Tailored Lagrangian relaxation forthe open pit block sequencing problem. Ann. Oper. Res. 222(1):419–438.
Lambert WB, Brickey A, Newman AM, Eurek K (2014) Open-pit block-sequencing formulations: A tutorial. Interfaces 44(2):127–142.
Lerchs H, Grossmann I (1965) Optimum design of open-pit mines. Cana-dian Mining Metallurgical Bull. 68:17–24.
Linderoth JT, Savelsbergh MWP (1999) A computational study of searchstrategies for mixed integer programming. INFORMS J. Comput.11(2):173–187.
Morton DP (1996) An enhanced decomposition algorithm for multi-stage stochastic hydroelectric scheduling. Ann. Oper. Res. 64(1):211–235.
Pochet Y, Wolsey L (2006) Production Planning by Mixed Integer Pro-gramming (Springer, New York).
Ramazan S (2007) The new fundamental tree algorithm for produc-tion scheduling of open pit mines. Eur. J. Oper. Res. 177(2):1153–1166.
Ramazan S, Dimitrakopoulos R (2007) Stochastic optimisation of long-term production scheduling for open pit mines with a new integerprogramming formulation. Orebody Modelling and Strategic MinePlanning (Australasian Institute of Mining and Metallurgy, Mel-bourne, Australia), 385–391.
Rojas CR, Goodwin GC, Seron MM, Zhang MM (2007) Open-cut mineplanning via closed-loop receding-horizon optimal control. SánchezPeña RS, Casín JQ, Cayuela VP, eds. Identification and Control(Springer, London), 43–62.
Rousseau L-M, Gendreau M, Feillet D (2007) Interior point stabilizationfor column generation. Oper. Res. Lett. 35(5):660–668.
Ruszczynski A (1986) A regularized decomposition method for mini-mizing a sum of polyhedral functions. Math. Programming 35(3):309–333.
Ryan DM, Foster BA (1981) An integer programming approach toscheduling. Wren A, ed. Comput. Scheduling of Public Transport
Urban Passenger Vehicle and Crew Scheduling (North-Holland,Amsterdam, Netherlands), 269–280.
Singh KJ, Philpott AB, Wood RK (2009) Dantzig-Wolfe decompositionfor solving multistage stochastic capacity-planning problems. Oper.Res. 57(5):1271–1286.
Stolarczyk L (1992) Definition imaging of an orebody with the radioimaging method. IEEE Trans. Indust. Appl. 28(5):1141–1147.
Van Den Heever SA, Grossmann IE (2000) An iterative aggregation/disaggregation approach for the solution of a mixed-integer nonlinearoilfield infrastructure planning model. Indust. Engrg. Chemistry Res.39(6):1955–1971.
Whittle J (1998) Four-X™ Strategic Planning Software for Open PitMines, Reference Manual (Whittle Programming Pty Ltd., Mel-bourne, Australia).
Wittrock RJ (1985) Dual nested decomposition of staircase linear pro-grams. Math. Programming Study 24:65–86.
Wollmer RD (1980) Two stage linear programming under uncertainty with0–1 integer first stage variables. Math. Programming 19(1):279–288.
Thomas W. M. Vossen is an associate professor in the Man-agement and Entrepreneurship division at Leeds School of Busi-ness, University of Colorado Boulder. His main research interestis the development and analysis of large-scale optimization meth-ods to support planning and scheduling decisions, in applicationareas ranging from cutting-stock problems and open-pit mining toair traffic flow management and network revenue management.
R. Kevin Wood is a Distinguished Professor Emeritus ofOperations Research at the Naval Postgraduate School (NPS).Retired as of 2015, he is continuing research on (a) mathematical-programming methods, (b) designing infrastructure that is robustagainst attack, and (c) skiing.
Alexandra M. Newman is a professor in the MechanicalEngineering Department at the Colorado School of Mines. Herresearch focuses on applied optimization, especially as it pertainsto mine production scheduling and energy systems.